ArticlePDF Available

Abstract and Figures

Most pictures shared online are accompanied by temporal metadata (i.e., the day and time they were taken), which makes it possible to associate an image content with real-world events. Maliciously manipulating this metadata can convey a distorted version of reality. In this work, we present the emerging problem of detecting timestamp manipulation. We propose an end-to-end approach to verify whether the purported time of capture of an outdoor image is consistent with its content and geographic location. We consider manipulations done in the hour and/or month of capture of a photograph. The central idea is the use of supervised consistency verification, in which we predict the probability that the image content, capture time, and geographical location are consistent. We also include a pair of auxiliary tasks, which can be used to explain the network decision. Our approach improves upon previous work on a large benchmark dataset, increasing the classification accuracy from 59.0% to 81.1%. We perform an ablation study that highlights the importance of various components of the method, showing what types of tampering are detectable using our approach. Finally, we demonstrate how the proposed method can be employed to estimate a possible time-of-capture in scenarios in which the timestamp is missing from the metadata.
Content may be subject to copyright.
Content-Aware Detection of
Temporal Metadata Manipulation
Rafael Padilha, Tawfiq Salem, Scott Workman, Fernanda A. Andal´
o, Anderson Rocha, and Nathan Jacobs
Abstract—Most pictures shared online are accompanied by
temporal metadata (i.e., the day and time they were taken),
which makes it possible to associate an image content with
real-world events. Maliciously manipulating this metadata can
convey a distorted version of reality. In this work, we present
the emerging problem of detecting timestamp manipulation. We
propose an end-to-end approach to verify whether the purported
time of capture of an outdoor image is consistent with its content
and geographic location. We consider manipulations done in the
hour and/or month of capture of a photograph. The central idea
is the use of supervised consistency verification, in which we
predict the probability that the image content, capture time,
and geographical location are consistent. We also include a pair
of auxiliary tasks, which can be used to explain the network
decision. Our approach improves upon previous work on a large
benchmark dataset, increasing the classification accuracy from
59.0% to 81.1%. We perform an ablation study that highlights the
importance of various components of the method, showing what
types of tampering are detectable using our approach. Finally,
we demonstrate how the proposed method can be employed to
estimate a possible time-of-capture in scenarios in which the
timestamp is missing from the metadata.
Index Terms—Timestamp verification, metadata manipulation
detection, digital forensics, temporal metadata manipulation.
WITH the popularization of social networks and ad-
vances in image capturing devices during the last
decade, the number of images shared online has grown ex-
ponentially. Their increasing availability coupled with easy-
to-use photo editing software has resulted in a profusion
of manipulated images. Such images are often maliciously
used to support false claims and opinions. Consequently,
the research community has explored many approaches to
detect image content tampering [1]–[3]. Even though visual
manipulation is now well-understood and reasonably explored
in the digital forensics literature, there is a more subtle type of
manipulation that still has been relatively unexplored by the
research community: timestamp manipulation.
Corresponding author: R. Padilha,
R. Padilha,F. A. Andal´
o, and A. Rocha are with the Institute of
Computing, University of Campinas, Brazil.
T. Salem is with the Department of Computer and Information Technology,
Purdue University, USA.
S. Workman is with DZYNE Technologies, USA.
N. Jacobs is with the Department of Computer Science, University of
Kentucky, USA.
This paper has supplementary downloadable material available at, provided by the author. The material includes a
list and description of the transient attributes and additional experiments
highlighting aspects of our method. Contact for
further questions about this work.
Fig. 1. Our goal is to analyze if an alleged timestamp is consistent with
the visual information present in a photograph. It is essential to consider
the appearance of the scene (e.g., illumination and weather conditions) and
photograph’s geolocation when verifying a timestamp.
Presenting an image as if it was captured in a different
moment in time can corroborate a false narrative and spread
misinformation. Recent examples include the “Fishwrap” cam-
paign [4], an online campaign that shared terror news articles
from past years in an attempt to spread fear and uncertainty.
In cases like this, articles often contain unaltered pictures that
are displaced in time to validate a false story. Even though the
timestamp of an image might be stored in its metadata when
the photo was originally taken, it can be easily tampered with
by freely available metadata editors (e.g., EXIF Date Changer
Lite or Metadata++).
In this sense, it is important to develop methods capable
of verifying the consistency between the visual information
of a picture and its timestamp. This is a challenging task that
requires a deep understanding of the scene, as its appearance
may vary depending on the hour, month, and the location
where the photo was taken (Figure 1). Moreover, factors such
as the weather, lighting conditions, device quality, and depicted
elements influence the appearance of a recorded scene and
directly affect our perception of time. Existing methods often
estimate indirect features from the image content, such as the
sun position in the sky [5]–[7] or meteorological measures [5],
and contrast them to registered values for the same day, hour,
and location. However, these are limited cues that may not
always be sufficient nor easily available.
We propose a convolutional neural network (CNN)-based
solution that analyzes a ground-level outdoor image and a
claimed timestamp (hour and month), receiving geographical
information as an additional context for its decision. The geo-
information can take the form of location coordinates (i.e.,
latitude and longitude) and/or basemap-style satellite imagery
associated with the location coordinates (but not necessarily
with the claimed timestamp). Our model is optimized in an
end-to-end manner using a multi-task loss function, with the
main goal of checking for manipulations and an auxiliary
objective of estimating transient attributes [8]. These are a
set of 40 attributes that encode the presence of high-level
properties of the scene appearance, such as weather (snow,
rainy,sunny), period of the day (sunrise,night,dusk), and even
subjective concepts (beautiful,stressful). This secondary task
introduces an explainability component to our model, aiding
in understanding its decisions.
We evaluate the proposed approach both quantitatively
and qualitatively, achieving state-of-the-art results on a
reference benchmark dataset [9]. Moreover, we perform
sensitivity analyses to understand how the appearance of the
scene, subtler timestamp manipulations, and noisy location
coordinates affect the verification performance of the method.
The contributions of our work include:
A new method for verifying the hour and month of cap-
ture of an outdoor image by comparing the appearance of
the scene with an alleged time and location information.
An extension of our method to estimate a possible time-
of-capture in scenarios in which the timestamp is missing
from the metadata.
A high-level exploration of which elements in a scene
might indicate timestamp inconsistency, as well as sev-
eral sensitivity analyses to help us understand when the
method works and fails.
A novel organization of the Cross-View Time dataset [9]
with camera-disjoint training and testing sets that are
more closely related to real-world scenarios.
The analysis of temporal information has been explored
in different ways in the literature. In this section, we review
relevant methods that approach this problem.
A. Metadata Tampering Detection
Traditionally, most tampering detection techniques focus on
image content manipulation, with only a few recent works
approaching metadata tampering detection. To check the in-
tegrity of the geographic location, researchers borrow ideas
extensively from the image retrieval and landmark recognition
literature [10]–[13], focusing on identifying visual elements
in the scene and matching them with large-scale databases of
geo-tagged images.
Considering timestamp manipulation, Kakar et al. [6] and
Li et al. [7] propose to verify the timestamp by estimating the
sun azimuth angle from shadow angles and sky appearance,
comparing it to the sun position calculated from the image
metadata. Their methods assume the camera to be perpendic-
ular to the ground and that the sky and at least one shadow
from vertical structures are visible in the scene. Chen et al. [5]
optimize a CNN to jointly estimate temperature, humidity, sun
altitude angle, and weather condition from an input image,
comparing them with meteorological data registered from the
day and time stored in its metadata. Even though they achieve
promising results, their approach requires access to historical
weather data that might not be available for most locations.
In [14], the authors approach this task as an event iden-
tification problem: they train a CNN with images from a
specific event (i.e., sharing similar location and timestamp)
to determine whether a test image, presumably sharing the
same metadata, belongs to that event. This assumption might
be valid considering forensic events with high media coverage,
but it might not be the case when most images are collected
from social media.
Differently from the mentioned works, we aim at directly
verifying whether a claimed time-of-capture is consistent with
the visual content of an outdoor image, independent of partic-
ular visual cues and without the need to optimize for a specific
event. With this in mind, we do not hold strong assumptions
on how the image was captured nor the appearance of the
In a similar line, the method proposed by Salem et al. [9]
learns a dynamic map of visual attributes that captures the
relationship between location, time, and the expected appear-
ance of a photograph. Their approach predicts visual attributes
from a combination of satellite imagery, time, and geographic
location. To apply the method to metadata verification, these
attributes are compared to similar information extracted from
a ground-level picture, computing a distance metric and using
it as a consistency score to detect if a timestamp has been
tampered with. In a similar manner, we train a global model
that captures how visual appearance relates to location and
time. However, we specifically focus on contrasting an alleged
timestamp against the visual characteristics of a picture by
directly optimizing for this task.
B. Time-of-capture Estimation
Approaching the problem from a different perspective, we
also analyze methods for directly estimating time-of-capture,
with time scales ranging from hours to decades. Several works
leverage specific visual elements as cues to date pictures, such
as human appearance and fashion [15], [16], visual style of
objects [17], [18], architecture styles [19], sun position [6],
[7], [20], and photo-generation artifacts [21]–[23].
Despite their impressive results, these methods require the
presence of particular visual elements in a scene to estimate
time reliably. In their absence, a different class of methods
should be considered. In this sense, a line of research closely
related to this work explores how the global appearance of a
scene changes over time.
Volokitin et al. [24] optimize a scene-specific classifier on
top of data-driven features to infer the time of the year and
hour of the day of images from a particular location. Some
works [8], [25] aim to estimate transient attributes of a scene
that inherently carry some degree of temporal information,
such as season and illumination conditions. By combining the
visual content with geographical information in the pretext
task of time-of-capture estimation, Zhai et al. [26] learn a
representation with high correlation to transient attributes.
This work builds upon these strategies, incorporating visual
information from a ground-level photo, geographical coordi-
nates to check for time-of-capture consistency, and optional
satellite imagery. Even though our goal is not to explicitly
estimate when an image was taken, we show how our method
can be applied for cases in which the timestamp is missing.
Our goal is to assess if the visual content of an outdoor
image is consistent with its hour and month of capture. For
this, our method must extract discriminative features from
the scene appearance and contrast them with the expected
appearance for that specific time of capture. As variations in
appearance over time are highly dependent on the location
of the scene, it is essential to provide, as additional context,
geographic cues of where the picture was taken [9], [26].
With this in mind, we propose a CNN architecture (Figure 2)
to estimate the probability P(y|G, t, l, S)that a given ground-
level image G, associated with location land satellite image
S, is consistent (y= 0) or inconsistent (y= 1) with an alleged
timestamp t. By providing location las input, the network will
be able to consider the influence of geographic position in
seasonal patterns (e.g., winter months in the Northern hemi-
sphere with reduced sunlight hours, and the opposite in the
Southern hemisphere). Moreover, satellite image Sprovides
an additional context about the photographer’s surroundings
and the structure of the scene, such as whether the image was
captured in an urban or rural area. We employ a basemap-
style picture, which is globally available and can be easily
obtained from online services (e.g., Google Maps and Bing
Maps) given location l. This kind of imagery was designed
for navigational purposes and offers an idea of the structure
of the scene without reflecting time-dependent elements (e.g.,
illumination and weather conditions). In this sense, we do not
assume Sis linked to the timestamp t, avoiding the need for
a satellite image at the precise time-of-capture being checked.
For explainability, the network also estimates transient at-
tributes aGand aS. These are 40-dimensional arrays, with
each value encoding the presence of a characteristic of the
scene appearance (e.g., fog, hot, beautiful, summer) to the
interval [0,1]. The attributes aGare estimated solely from the
ground-level image and capture the high-level properties of
the scene at the moment it was recorded. In contrast, aSis
estimated from the satellite photo, location coordinates, and
the alleged timestamp, and can be interpreted as a prediction
of the expected scene appearance at the alleged moment.
A. Network Architecture
Each input G, t, l, S is processed by individual sub-
networks that extract characteristics from each modality and
encode them into 128-dimensional feature vectors.
Visual Encoder. Both the ground-level and satellite images,
Gand S, are processed by a backbone CNN, extracting
feature maps from the last convolutional layer. The feature
maps are then processed by two additional fully-connected
layers, with 256 and 128 units, respectively, each followed
by a ReLU activation and batch normalization, resulting in
feature vectors fGand fS.
Location and Time Encoder. We represent location lin
earth-centered earth-fixed (ECEF) coordinates, scaled to
[1,1] by dividing each one by the radius of the Earth;
whereas time tis represented by the month and hour of
the day (UTC) individually scaled to [1,1] (i.e., 0 AM of
January would lead to (1,1) and 11 PM of December
to (1,1) coordinates). Both sub-networks have similar
architectures and comprise three fully-connected layers, with
256, 512, and 128 neurons, respectively, each followed by a
ReLU activation function and batch normalization, resulting
in feature vectors fland ft.
Task-specific Branches. Once each feature fG,fS,fl, and ft
has been extracted, they are fed to task-specific branches. Each
branch consists of three fully-connected layers. The initial two
layers have 256 and 512 units, respectively, followed by ReLU
and batch normalization, while the third maps the intermediate
features to the output dimension of each task. The top branch
receives the concatenation of {fG,fS,fl,ft}and outputs
the consistency y. Its last fully-connected layer has two units
followed by a softmax operation, allowing us to interpret the
output as the probability P(y|G, t, l, S). The middle branch
receives as input fGand outputs a set of transient attributes
aG. Likewise, the bottom branch processes the concatenation
of {fS, fl, ft}and outputs attributes aS. Their final dense layers
have a number of neurons matching the quantity of estimated
transient attributes |ˆa|(in our experiments, |ˆa|= 40), followed
by a sigmoid activation.
B. Loss Function
Our network is optimized in an end-to-end fashion both for
timestamp consistency and transient attribute estimation tasks.
Even though the main goal is timestamp tampering detection,
the auxiliary tasks allow our model to predict properties of
the scene, separately considering the alleged timestamp or the
ground-level visual information. Both sets of attributes can be
compared, offering insights into the model decision.
For the consistency verification branch, we calculate the
binary cross-entropy loss
LCE (y, ˆy) = ˆylog(y)(1 ˆy) log(1 y)(1)
between the consistency prediction yand the ground truth ˆy.
For the transient attribute branches, we compute the mean
squared error between estimated transient attributes aand
ground truth ˆa:
LMSE(a, ˆa) = 1
We compute two separate LMSE terms with respect to aGand
aS. By doing so, the model learns to extract the transient
attributes directly from the ground-level image while also
being able to estimate them from the satellite photo, location
Fig. 2. An overview of our approach. A ground-level image G, a timestamp t, geo-coordinates l, and a satellite image Shave their features extracted by
encoder networks and then fed to task-specific branches. One network branch predicts the consistency label y, while auxiliary branches estimate transient
attributes aGand aS. The network is optimized by a combination of cross-entropy and mean squared error losses. At inference time, the consistency answer
yis considered for the tampering detection, while aGand aSoffer insights about the decision of the network.
coordinates, and given timestamp. Finally, the whole network
is jointly optimized to minimize their weighted sum:
L=αLCE (y, ˆy) + βLMSE (aG,ˆa) + γLMSE(aS,ˆa).(3)
Even though setting different weights (α,βand γ) to each
loss term in Equation (3) might allow the network to focus on
a particular task and improve the detection performance, in
our experiments, we found that equal importance to all terms
achieved better results. We perform our experimental analysis
in Section IV with this design decision.
We evaluated our method for tampering detection, in which
it outputs whether the timestamp is consistent with the visual
attributes of the ground-level image. We assessed the quality
of our approach considering the accuracy (Acc) and the re-
ceiver operating characteristic (ROC) curves in the tampering
detection. We performed an ablation study of different input
modalities and the backbone CNNs used as visual encoders,
comparing them to existing approaches from the literature.
We also evaluated its sensitivity to realistic conditions, with
changes in the appearance of the scene, subtler timestamp
manipulations, and noisy location information. Additionally,
for scenarios in which the timestamp of an image is missing,
we demonstrated how our method could be applied to estimate
a possible moment of capture. Finally, we interpret the deci-
sions of our approach by analyzing image regions with strong
influence on the model’s predictions and mismatches between
estimated transient attributes.
A. Dataset and Training Details
We adopted the Cross-View Time dataset [9], comprising
more than 300k ground-level outdoor images. The dataset
combines 98k images from 50 static outdoor webcams of
the Archive of Many Outdoor Scenes (AMOS) [27], and
206k geo-tagged smartphone pictures from the Yahoo Flickr
Creative Commons 100 Million Dataset [28]. Ground-level
images were captured worldwide in different hours of the
day and months throughout the year, and are associated with
geographical coordinates and a timestamp (UTC). Given the
latitude and longitude of ground-level pictures, the authors
also collected co-located satellite imagery, downloading from
Bing Maps orthorectified overhead image centered on the
geographic location. Images have dimensions of 800 ×800
pixels and a spatial resolution of 0.60 meters/pixel. These are
not captured at the same moment of the timestamp, but instead
are basemap-style imagery intended to provide a clear picture
of the current configuration of buildings and roads. This type
of photograph offers plenty of details about the surrounding of
the photographer without reflecting weather conditions at the
time of capture (e.g., clouds, snow, rain). Lastly, the authors
provided for each ground-level image the set of 40 transient
attributes extracted with the method from [8] and encoded to
the interval [0,1].
We employed the same data splits provided by the authors,
training on 280k images and testing on 25k. The original
splits share locations (i.e., the same place might be represented
through different images in training and testing) and cameras,
as imagery from AMOS has been randomly sampled for
each set. In Section V, we propose a novel and challenging
organization of the dataset that considers camera-disjoint sets.
For training, batches were randomly sampled from the train-
ing images, with corresponding timestamps, geo-coordinates,
and satellite images. For each consistent sample, we generated
a tampered version by exchanging its timestamp to that of a
random image in the training set. Due to the hour distribution
of the set — which mostly comes from social media and,
thus, is concentrated on the 10 AM to 8 PM interval in which
people take more photos — it is less likely that the majority of
images will completely change their period of the day after the
manipulation. This type of tampering creates a more realistic
scenario for our evaluation than randomly sampling a tampered
timestamp throughout the day and year, as done by previous
works [7]. Considering this, training batches were composed
of the same number of consistent and inconsistent tuples. At
test time, a similar process was used, generating a tampered
tuple for each available test image.1
We use a VGG-16 [29] network pre-trained on the Places
dataset [30] as the feature extractor for the ground-level image,
while a ResNet-50 [31] network pre-trained on ImageNet [32]
processes the satellite image. This is a similar architecture
to [9] which allows a fair comparison to their method. Before
being processed by the networks, we resize Gand Sto 224 ×
224 and scale each pixel to [1,1].
Our architecture was optimized using Adam [33] with an
initial learning rate of 105, batches of 32 images, and trained
for 30 epochs. When calculating the loss for a batch, only
real tuples were considered in computing the LMSE terms
of Equation (3), as the ground-truth transient attributes for
tampered timestamps are not available. With the exception of
VGG-16 and ResNet-50 sub-networks, the weights of convolu-
tional and fully-connected layers were initialized with Xavier
initialization, and we applied L2regularization (λ= 0.001).
B. Ablation Study
Each input modality influences the performance of the
timestamp tampering detection. We performed an ablation
study, considering the impact of location land satellite image
S, given as additional context to the ground-level image
Gand timestamp t. To better evaluate the impact of each
modality, we optimized the models considering solely the
consistency verification, i.e., by removing the mean-squared
error terms from Equation (3). We also considered variations
of our architecture in which we use a ResNet-50 [31] or
a DenseNet-121 [34] as encoders for both ground-level and
satellite images. In this scenario, even though the branches
are similar, they do not share weights as each processes a
different type of input. Finally, considering all input modalities
and evaluated backbones, we optimized the model with the
transient attribute estimation task by employing the complete
loss function from Equation (3). We present the accuracy
and the area under the receiver operating characteristic curves
(AUC) for our evaluation in Table I.
Our results show that including geographical coordinates
(G, t, l) considerably improved performance, as the model
learns how shifts in latitude and longitude correlate to the
expected appearance of a scene and an alleged timestamp
(e.g., a snowy scene in December might be consistent if
taken in the Northern hemisphere, but less so in the South-
ern hemisphere). The inclusion of satellite imagery (G, t, S)
also boosts performance, but the gain is slightly inferior to
that of the location-only model (G, t, l). In addition, when
1The same seed was selected for all experiments and
the tampered timestamps of test images are available in
VGG-16 / ResNet-50 ResNet-50 DenseNet-121
Modalities Acc (%) AUC Acc (%) AUC Acc (%) AUC
G, t 63.6.699 65.3.744 67.5.766
G, t, l 77.9 .853 75.1.852 78.7.873
G, t, S 72.1.813 74.9.851 77.0.855
G, t, l, S 76.1.847 78.4.877 80.5.880
G, t, l, S (TA) 75.7.834 78.7 .865 81.1 .885
Salem et al. [9] 59.0.627 – – – –
Volokitin et al. [24] 51.7.517 – – – –
comparing to each individually (G, t, S and G, t, l models),
the simultaneous inclusion of coordinates and satellite data
improves the results even further (G, t, l, S). This indicates
that both input modalities share complementary information
that is being captured by the model. Even though one might
think the aerial modality bears a collection overhead that
might not be worth considering its performance gain, satellite
data is globally available and can be directly obtained from
location coordinates. In this sense, it would be more likely
not to have any sort of geographical information (e.g., if the
metadata is completely erased) than not being able to collect
an aerial image. Nonetheless, the location-only model (G, t, l)
still achieved competitive results and could be employed in
real application scenarios as well.
Replacing the backbone CNNs to ResNet-50 or DenseNet-
121 improved the results even further. Even though VGG-16 is
frequently used for transfer learning, the other architectures in-
clude several operations—such as residual learning and batch
normalization—that boost optimization and performance.
Finally, the addition of the auxiliary optimization tasks—
G,t,l,S(TA) in Table I—slightly improved the results for
the ResNet-50 and DenseNet-121 setups. Even though the
estimated attributes, aGand aS, are not considered when de-
termining if tampering occurred, in Section IV-H we use them
to produce a simple explanation of the network’s decision.
C. Comparison to Existing Approaches
We compared our models with the method proposed by
Salem et al. [9], using the same weights and choice of
hyperparameters provided by the authors. They employ a
similar architecture to ours—with VGG-16 and ResNet-50
as feature extractors for the ground-level and satellite im-
agery, respectively—that extracts features related to scene
classification and transient attributes. In a similar fashion,
one set of features is obtained directly from the ground-level
picture and another from the satellite image, location, and
timestamp. They compute a distance between both sets (based
on KL divergence and L2), using it as a consistency score for
timestamp tampering detection.
Even though the authors apply their approach for timestamp
verification, the learning process focuses on capturing the
geographical and temporal behavior of transient attributes.
Though this may aid in learning richer features, as shown
by our ablation study, they are not discriminative enough by
(a) Same location recorded in April under different hours of the day. (b) Same location recorded at 6P M in different months of the year.
Fig. 3. Consistency probability for a scene recorded in different moments in time. Each curve represents a fixed alleged timestamp being verified against
ground-level images captured in different (a) hours of the day in April and (b) months of the year at 6PM. Our model captures several temporal patterns,
identifying how the appearance of the scene changes according to the period of the day and season.
themselves for verifying time-of-capture. Our results show that
optimizing the network specifically for this task was essential
to better detect manipulations, outperforming their approach
by 22 percentage points in accuracy.
Additionally, we evaluate the method of Volokitin et al. [24]
for this task. We trained Random Forest classifiers on top of
features extracted with a pre-trained VGG model to estimate
the month and hour of capture of an image. To apply their
approach to the verification scenario, we computed the class
probability of month and hour classifiers for each test image,
multiplied them, and compared against a threshold selected in
a held-out validation set.
Differently from ours, their method does not leverage ge-
ographic information but instead trains each classifier with
images from a single location. This might be viable in
particular cases in which there are enough images to train
a location-specific verification model but becomes infeasible
as we analyze geographically-spread imagery, such as the
CVT dataset. Moreover, optimizing for verification—instead
of adapting a timestamp estimation method—allows our model
to capture contrasting information between alleged timestamp
and scene appearance essential to detect manipulations.
Finally, most works mentioned in Section II have strong
assumptions about the input data that do not hold valid for
images originating from social media. In such a scenario, avail-
able data is collected under unconstrained capturing conditions
(i.e., presenting varied illumination and weather conditions,
from different points of view and angles, having varied quality
and resolution). Even though methods that rely on identifying
particular visual clues (e.g., shadows from vertical structures,
sun position in the sky, architectural or object style, fashion)
might achieve accurate results when their underlying assump-
tions are met, they are often inadequate for general situations
such as the one we tackle.
D. Sensitivity Analysis: Scene Appearance
We explored the sensitivity of our method to changes in the
appearance of a scene under fixed alleged timestamps. In a first
evaluation, we selected images from different hours of the day,
all taken in a particular month at the same location. Whereas,
in a second experiment, we employed images taken at the
same hour and location but in different months. We computed
the consistency probability of each image under fixed alleged
timestamps, and plotted the resulting curves in Figure 3. By
doing so, we investigate how confident the network is that an
alleged timestamp is consistent as time progresses in a scene.
The model correctly predicts high probabilities for images
taken around the alleged timestamp, as seen by the peaks in the
curves from Figure 3(a). Despite that, nighttime images tend
to be very similar due to the lack of changes in illumination,
reflecting in higher consistency probabilities between 9PM and
4AM. Similarly, Figure 3(b) shows that our model captured
seasonal patterns, such as the variation in sunset hours across
the year, even though monthly variations in the appearance of
a scene tend to be smoother across neighboring months. We
present in the Supplementary Material additional examples and
an evaluation considering multiple cameras from AMOS.
E. Sensitivity Analysis: Timestamp Manipulation
In real scenarios, when the timestamp of a photograph
is modified, the new time-of-capture is typically selected
considering a plausible, often close to the original, moment
in time as a way to make the detection harder. For example,
claiming a 9AM image was captured in daylight hours is more
convincing than saying it was captured at night.
To evaluate this scenario, we performed several experiments
in which all images in our test set were tampered with by an
equal month and hour shift in both directions to its original
time-of-capture. For a given shift pair (tmonth,thour), we
tampered with every testing image by adding ±tmonth and
±thour to its time of capture, effectively producing four tam-
pered versions for each image. For example, for tmonth = 1
and thour = 2, a picture captured in (Dec, 11AM) would
generate timestamps (Jan, 9AM), (Jan, 1P M), (Nov, 9AM)
and (Nov, 1PM). We evaluated the detection rate of such
manipulated images and present the results in Figure 4(a).
As we expected, the closer an alleged timestamp is to
the original moment of capture, the harder it is to detect
that tampering occurred. Shifts by a single hour or month
were detected in less than 20% of the cases. Due to the
similarity in the appearance of the scene, these are complex
to detect solely based on pixel values and without any other
(a) Model trained with random manipulations. (b) Model fine-tuned for subtler manipulations.
Fig. 4. Detection rate for hour and month shifts from the original timestamp for (a) model trained with randomly sampled timestamp manipulations and (b)
model fine-tuned for 10 additional epochs sampling manipulations close to the ground-truth timestamp. Each curve represents a month shift (tmonth), while
the x-axis denotes different hour shifts (thour). By fine-tuning it with harder-to-detect tampering, the model learns to identify subtler manipulations of a few
hours and months (bottom left region of the plots), improving the detection rate of such cases.
hard assumptions on the images. This becomes even more
challenging considering that modern cameras make use of
automatic exposure adjustment techniques that compensate for
brightness distortions during capture, making a scene darker
or brighter than it is in reality. Besides obliterating image
differences caused by 1-hour time spans, artificial changes in
the brightness of the picture might be perceived as an alteration
caused by the progression of time, thus influencing the model
decision. Existing works able to detect such manipulations
more accurately either consider weather and astronomical
information [5] or rely on several combined images of the
same time and place [14], which are not always available for
images originated from social media.
As the gap between ground truth and alleged timestamp
increases, the inconsistency between time and appearance
becomes more apparent and detectable. E.g., images that were
presented as being captured in a different season (tmonth
3) or period of the day (thour 6, such as claiming a
morning scene was captured at noon or at night) were detected
in more than 75% of the cases.
As we trained our models with random sets of tampered
month and hour, in accordance with the evaluation protocol
from [9], the network learns how to better capture clearer
manipulations, as they tend to be sampled more frequently
in comparison to sets closer to the ground-truth timestamp.
However, it is possible to shift the focus to detecting subtler
and harder manipulations as well. For that, we fine-tuned our
best model (DenseNet — G, t, l, S - TA) for 10 epochs
using tampered timestamps closer to the ground-truth time-of-
capture. Each manipulated timestamp was generated randomly
selecting tmonth and thour in 1,±2}. We present the
results in Figure 4(b). The detection rate for such cases
improves considerably as the model now captures the contrast
between neighboring hours and months. In the general/random
Proposed Approach Location Augmentation Salem et al. [9]
lAcc (%) AUC Acc (%) AUC Acc (%) AUC
Without noise 81.1.885 75.9.848 59.0.627
166.3.718 75.9.848 52.2.539
566.3.721 76.0.847 52.1.538
1565.9.716 75.8.846 51.9.533
3064.3.695 74.8.838 51.6.529
4562.2.686 73.5.824 51.6.528
6061.1.677 72.5.811 51.6.528
7560.5.671 71.7.801 51.6.527
case, this model has a minor decrease in accuracy (77.78%)
while slightly improving AUC (0.888).
F. Sensitivity Analysis: Geographic Location
As shown in our ablation study, the geographic coordinates
and satellite image provide an essential context to assess the
consistency between an alleged timestamp and the ground-
level photograph. However, in real cases, the precise location
in which an image was captured might not be known, and only
a rough set of coordinates might be available instead. In this
sense, we evaluate the impact of noisy location coordinates
on the performance of our approach. For each sample in our
testing set, we add ±lto its original latitude or longitude,
artificially altering the location of that image. As the satellite
image is directly related to the geographic coordinates, we also
collect a new aerial photograph for the perturbed position. We
evaluate the accuracy and AUC under different values of l
for our best ablation model and the method of Salem et al. [9],
and present the results in Table II.
Fig. 5. Heatmap of the consistency probability distribution over local time (all possible months and hours of capture) for proposed method trained with
randomly sampled manipulations and alleged timestamps close to the ground truth, in comparison to Salem et al. [9]. For each example, we show the ground-
level picture and satellite image of that location. Yellow areas represent high consistency, and blue indicates inconsistent moments in time. The red dot marks
the ground-truth moment-of-capture. Note how the proposed methods provide high-consistency month/hour pairs close to the ground-truth timestamp with
fewer false activation regions in comparison to Salem et al. [9]. We provide additional examples in the Supplementary Material.
Feeding noisy sets of coordinates and satellite photograph to
the models, as expected, negatively impact their performance.
This highlights the sensitivity of the models when we assume
correct locations but evaluate it under a noisy scenario. The
performance of our approach is stable when considering
perturbations smaller than 15. Each longitudinal movement
of 15represents a one-hour shift, which is the smallest
manipulation captured by the method. In this sense, minor
degrees of noise do not affect the performance of the approach.
However, when the noisy coordinates deviate considerably
from the ground-truth location, the model interprets these dif-
ferences as signs of manipulation. Most error cases are related
to consistent pairs of ground-level image and timestamp that
are mistakenly classified as manipulations due to inconsistent
coordinates. It also might indicate that the model is over-
reliant on the coupled information of ground-level picture and
geographic input data, and visual inconsistencies that might
be created by combining a photograph with a satellite image
from another location (e.g., an urban scene paired with a rural
or coastal aerial picture) are identified as a sign of tampering.
This is especially true for samples from the Flickr subset of
CVT, as each geographic location might be represented by a
single picture. We present in the Supplementary Material an
extension of this experiment, evaluating more degrees of noise
and discriminating the results per subset of the CVT dataset.
Considering this, to improve robustness to unreliable loca-
tion information, we retrain our model with randomly per-
turbed geographic coordinates. In each training batch, we
jitter the location coordinates of 50% of batch samples by a
noise ±l∈ {0.05,0.1,0,25,0.5,1,5,10,15}. The
training is performed as detailed in Section IV-A and we
include the results in Table II.
The Location Augmentation strategy outperformed the other
methods when we consider noisy geographic information.
The impact of unreliable location on it is less noticeable
than that observed in our model optimized with ground-truth
geographic information and the approach from Salem et al [9].
This highlights the importance of introducing techniques that
improve the robustness to location errors, especially if we
want to apply these models in unconstrained scenarios, such
as when verifying footage from social media.
G. Time Estimation: A Qualitative Exploration
We also investigated the application of our method when
a timestamp is not available. For a ground-level photo with
associated location and satellite image, we predicted the con-
sistency probability P(y|G, ti, l, S),tiT, with Tbeing
all combinations of month and hour of capture.
In Figure 5, we show the consistency distribution heatmap
for a few examples, as well as their ground-truth times-
tamp. We also include the consistency heatmaps produced by
our model fine-tuned to detect subtler manipulations (Sec-
tion IV-E). In comparison, we generate heatmaps with the
Fig. 6. Occlusion activation maps [35] for two pictures under different
timestamps. The network focuses on specific elements depending on the
alleged timestamp, with yellow regions representing important elements and
blue areas having a low impact on the decision. Activation tends to be
spread homogeneously throughout the image, as the network focuses on the
illumination and overall appearance of the scene. However, the model is also
able to shift its attention to clear telltales, such as city lights during the day.
We provide additional examples in the Supplementary Material.
method from Salem et al. [9], computing the consistency
score for all possible timestamps in a similar manner to
Section IV-B.
Even though our method is not explicitly trained to predict
time, it is able to coherently estimate a possible span in
which an image might have been captured. As it properly
learned the influence of time in the appearance of a scene,
capturing intrinsic temporal patterns, it also produces more
precise estimations than [9]. The model optimized to detect
subtler manipulations is even better at assigning areas of high
consistency closer to the ground-truth timestamp with less
false activation on other month/hour pairs.
H. Summer Snow and Midnight Sun: Tampering Telltales
As humans, we have a good intuition of which elements in
a scene might be inconsistent with an alleged time of capture.
For example, we can easily spot the inconsistency of a bright
sky in a nighttime picture or of snow-covered ground in a
summer scene. Besides achieving high accuracy in tampering
detection, we want to provide direct explanations of what
elements in the scene might be the cause of inconsistency
according to our models. To this end, we investigated which
image regions have the most impact on the decision of
the network for both consistent and inconsistent examples.
Similarly to [35], we occluded, in a sliding window manner,
parts of the input image with gray patches of sizes 50 ×50,
100×100,100 ×50 and 50 ×100 and evaluated the difference
in consistency probability y. When an important area of a con-
sistent image is occluded, we expect yto decrease; whereas,
occlusions over an inconsistent sample might increase y, as it
might hide discrepant elements. Figure 6 depicts the occlusion
maps for two ground-level images, considering the ground-
truth and manipulated timestamps. Additional examples are
presented in the Supplementary Material.
The network focuses on specific elements depending on the
time of capture being examined. When claiming the picture
was captured in a different moment of the day, the model
Fig. 7. Comparison between a subset of transient attributes aGand aSfor
two scenes, for consistent (top) and inconsistent (bottom) timestamps. Our
model correctly matches both sets of attributes for consistent timestamps.
However, as it estimates aSwithout using ground-level signal, these attributes
substantially differ from aGfor inconsistent time-of-captures. We provide
additional examples in the Supplementary Material.
often considers regions spread across the image, as the scene
changes uniformly with the shift in illumination. Besides that,
it also activates for some inconsistent elements, such as shad-
ows (bottom-left corner of top example) and nighttime city
lights (bottom example). Similarly, when tampering with their
month, it considers background vegetation and lights reflected
by the water, in the top and bottom scenes, respectively.
Besides the occlusion maps, the transient attributes (aG
and aS) could provide additional evidence for the network
decision. Even though they do not significantly improve
the detection performance, as shown in our ablation study
(Section IV-B), they encode how our model perceived the
appearance of the scene based on two distinct sets of inputs.
As aGis obtained strictly from the ground-level image and
is not influenced by a possibly manipulated timestamp, it
captures the characteristics of the scene when the picture was
taken. Differently, aSis predicted from the satellite photo,
location coordinates, and the alleged timestamp, predicting
the expected scene appearance at that alleged moment. In
this sense, in case the timestamp tmatches the ground-
truth moment-of-capture, we expect aGand aSto be similar,
whereas an inconsistent tmight lead to discrepancies in both
sets of attributes. The analysis of such inconsistencies is useful
for explainability purposes.
We compare the sets of transient attributes for the same loca-
tion taken at different moments. For each scene, we extracted
attributes aSconsidering two timestamps, one matching the
ground-truth time-of-capture and another from a tampered
timestamp. We present in Figure 7 their comparison against
aG. To highlight their differences, we selected the top five
divergent attributes from the inconsistent setup.
When considering the same moment in time, aGand aStend
to be similar, while an inconsistent timestamp will produce
discrepant transient attributes. For the inconsistent example
on the left column of Figure 7, our model expected it to
be a sunny day of summer for the (August, 7AM) time-of-
capture, in contrast to the nighttime scene in the ground-level
image (December, 2AM). We provide more examples in the
Supplementary Material.
I. Transient Attribute Influence Analysis
The distribution of transient attributes estimated by our net-
work reflects some aspects of scene appearance that might be
inconsistent with the alleged timestamp. As shown in Figure 7,
most inconsistent attributes are related to illumination and
seasonal patterns, which might indicate they are discriminative
enough to verify a timestamp.
Figure 8 presents an analysis of how the network decision
is reflected into the set of estimated transient attributes. For
each transient attribute in aGand aS, we computed a joint
histogram over their predicted values for consistent and in-
consistent examples in our testing set that were classified with
more than 90% confidence. As expected, estimated attributes
are similar for consistent examples, with a high count located
in the diagonal of the histograms. Whereas for inconsistent
examples, the divergence between aGand aSpredictions is
highlighted by activation spread outside the diagonal.
For every histogram, we computed the mutual information
(MI) considering the distributions obtained by aGand aS.
This measures the dependency between an attribute estimated
from the ground-level image or from satellite, location, and
timestamp. In this sense, the MI calculated from consistent
examples tends to be higher due to the similarity between aG
and aSthan that of inconsistent examples. Considering this,
we ranked each transient attribute based on the difference in
MI from consistent and inconsistent classes. This quantifies
how discriminative an attribute is to capture divergences in aG
and aSdue to a tampered timestamp. Figures 8(a) and 8(b)
show the top and bottom four ranked attributes, respectively.
Most discriminative attributes are related to the day-night
cycle and changes in the illumination of the scene, with varied
histogram distribution. Complementarily, lush captures the
vegetation of scenes, an important element across months and
seasons. On the other hand, attributes such as busy and sen-
timental show similar histograms between both classes, while
rain and storm are concentrated on a few bins, which might
reflect the low occurrence of images with such conditions in
our dataset.
We also present, for these eight attributes, the normalized
histogram of |aGaS|for both classes. For discriminative
attributes, the distributions are considerably different, with
consistent examples concentrated on lower bins and inconsis-
tent ones spread towards higher differences between attributes.
Differently, less informative attributes show similar histogram
distribution for both consistent and inconsistent examples. We
present in the Supplementary Material additional experiments
comparing the estimated set of attributes for both classes.
The standard evaluation protocol of Cross-View Time
dataset [9] allows for certain cameras to be shared between
training and testing sets. This protocol can emulate scenarios
in which we need to verify the authenticity of images from
a particular set of devices and locations. Considering the
ubiquity of surveillance systems (CCTV) nowadays, this is a
common scenario, especially for big cities and high visibility
events (e.g., protests, musical concerts, terrorist attempts,
sports events). In such cases, we can leverage the availability
of historical photographs of that device and collect additional
images from previous days, months, and years. This would
allow the model to better capture the particularities of how
time influences the appearance of that specific place, probably
leading to a better verification accuracy. However, there might
be cases in which data is originated from heterogeneous
sources, such as social media. In this sense, it is essential
that models are optimized on camera-disjoint sets to avoid
learning sensor-specific characteristics that might not general-
ize accordingly for new imagery during inference.
With this in mind, we propose a novel organization for CVT
dataset.2We split available data into training and testing sets,
ensuring that all images from a single camera are assigned
to the same set. During this division, we aimed to keep the
size of each set roughly similar to the original splits, allowing
models to be optimized with similar amounts of data.
Under this cross-camera protocol, we retrained our best
model (DenseNet – G,t,l,S, TA), achieving an accuracy
and AUC of 67.9% and .749, respectively. When we consider
the model optimized without the satellite image (DenseNet
G,t,l), it achieves 65.7% and .735 on the same metrics.
Even though the aerial image improves performance similarly
to our previous experiments, the location-only model is still
competitive. In comparison, we trained the method from Salem
et al. [9], using code provided by the authors, and obtained
an accuracy of 55.0% and AUC of .565. Both proposed
models outperformed by a large margin the technique from [9],
highlighting the importance of combining the information
from multiple input modalities and training specifically for
the timestamp verification task.
Despite that, we see a performance drop for all methods in
comparison to the previous experiments. This was expected
due to the more challenging conditions of the new protocol.
As there will be locations not covered in the training set,
for such cases, the network must take into account in its
decision the knowledge obtained from other training samples
roughly from that latitude and longitude interval. Besides that,
under this new protocol, the models are required to generalize
to spatially-distributed sources without relying on camera-
specific cues. This is particularly important considering that
roughly a third of CVT dataset is originated from 50 static out-
door webcams of the Archive of Many Outdoor Scenes [27]. In
a scenario with shared devices between sets, the model could
focus on learning patterns specific to these cameras, which
allows it to achieve superior accuracy at inference but hinders
its generalization to unseen places and devices.
Nonetheless, as more data becomes available, covering more
locations and timestamps, we expect the performance on both
camera-disjoint and shared-camera setups to be similar.
We introduced a novel approach for detecting if the alleged
capture time (hour and month) of an outdoor image has
been manipulated. Our architecture incorporates inputs from
2The data organization of the camera-disjoint sets is available in
(a) Top-ranked attributes. (b) Bottom-ranked attributes.
Fig. 8. Joint histograms for eight transient attributes estimated from ground-level picture (aG) or satellite image, location and timestamp (aS) for consistent
and inconsistent examples in our test set. Attributes were ranked based on the difference in mutual information between the joint histograms of both classes,
highlighting (a) discriminative and (b) less informative attributes. We also show the histogram over |aGaS|, emphasizing the difference in distributions of
each class.
location, time, and satellite imagery, and is jointly optimized to
detect timestamp tampering and estimate high-level attributes
about the scene appearance. The proposed approach achieved
high detection accuracy, improving the state-of-the-art on a
large-scale dataset and surpassing existing approaches while
having fewer assumptions and more realistic applicability.
We demonstrated that incorporating geographical context —
in the form of location coordinates and/or satellite imagery —
was essential for this problem, allowing the model to account
for geographic patterns that influence scene appearance in
a particular month and hour. Just like location and satellite
imagery, our method can be easily adapted to receive other
types of data as additional context. High-quality information,
such as meteorological data similar to [5] or scene classifi-
cation [30], could complement present features and improve
tampering detection.
While optimizing the model to estimate transient attributes
slightly improved the detection accuracy, their major benefit
was the added interpretability factor. They allow us to generate
explanations for the classification decisions by comparing a
set of attributes estimated solely from the ground-level image
against another set predicted from the timestamp and location
modalities. Mismatching attributes often indicate evidence of
tampering captured by our model. Complementarily, we also
showed how to interpret an outcome by analyzing pixel-level
influence on the model’s decisions.
Additionally, we analyzed the sensitivity of our approach
under realistic scenarios, with unreliable location information,
subtler timestamp manipulation, and changes in the appearance
of the scene. Even though most of them influence the model’s
ability to detect manipulations, their impact can be undermined
by applying data augmentation techniques during training.
In particular, sampling manipulated timestamps closer to the
ground-truth moment-of-capture of images, as well as slightly
jittering location coordinates of training samples improved the
model robustness to such cases.
We also demonstrated how our method could estimate a
possible time of capture for an image missing its timestamp.
By shifting how we address the problem, from a verification to
an estimation scenario, we show that the proposed method cor-
rectly learned discriminative temporal patterns, outperforming
existing approaches in this task.
Finally, we propose a new cross-camera evaluation protocol
for the Cross-View Time dataset [9], with camera-disjoint
training and testing sets to emulate a realistic application
scenario. Under this protocol, trained models are required to
generalize for geographically-spread imagery without relying
on location- or sensor-specific cues.
The authors would like to thank the S˜
ao Paulo Re-
search Foundation (FAPESP, grants #2017/21957-2 and
#2019/15822-2) and the US National Science Foundation (IIS-
1553116) for the financial support.
[1] V. Cristlein, C. Riess, J. Jordan, C. Riess, and E. Angelopoulou, “An
evaluation of popular copy-move forgery detection approaches,” IEEE
Trans. Inf. Forensics Security, vol. 7, no. 6, pp. 1841–1854, 2012.
[2] P. Zhou, X. Han, V. I. Morariu, and L. S. Davis, “Learning rich features
for image manipulation detection,” in IEEE Conf. on Computer Vision
and Pattern Recognition, 2018, pp. 1053–1061.
[3] M. Huh, A. Liu, A. Owens, and A. A. Efros, “Fighting fake news:
Image splice detection via learned self-consistency,” in European Conf.
on Computer Vision, 2018, pp. 101–117.
[4] A. Hern, “’fishwrap’ fake news campaign recycles old news of terror
attacks,” in The Guardian. Online;, 2019.
[5] B.-C. Chen, P. Ghosh, V. I. Morariu, and L. S. Davis, “Detection of
metadata tampering through discrepancy between image content and
metadata using multi-task deep learning,” in IEEE Conf. on Computer
Vision and Pattern Recognition Workshops, 2017, pp. 60–68.
[6] P. Kakar and N. Sudha, “Verifying temporal data in geotagged images
via sun azimuth estimation,” IEEE Trans. Inf. Forensics Security, vol. 7,
no. 3, pp. 1029–1039, 2012.
[7] X. Li, W. Xu, S. Wang, and X. Qu, “Are you lying: Validating the time-
location of outdoor images,” in Intl. Conf. on Applied Cryptography and
Network Security, 2017, pp. 103–123.
[8] P.-Y. Laffont, Z. Ren, X. Tao, C. Qian, and J. Hays, “Transient attributes
for high-level understanding and editing of outdoor scenes,ACM Trans.
on Graphics, vol. 33, no. 4, p. 149, 2014.
[9] T. Salem, S. Workman, and N. Jacobs, “Learning a dynamic map of
visual appearance,” in IEEE Conf. on Computer Vision and Pattern
Recognition, 2020.
[10] J. Hays and A. A. Efros, “IM2GPS: estimating geographic information
from a single image,” in IEEE Conf. on Computer Vision and Pattern
Recognition, 2008, pp. 1–8.
[11] H. Noh, A. Araujo, J. Sim, T. Weyand, and B. Han, “Large-scale image
retrieval with attentive deep local features,” in IEEE Intl. Conf. on
Computer Vision, 2017, pp. 3456–3465.
[12] F. Radenovi´
c, G. Tolias, and O. Chum, “Fine-tuning cnn image retrieval
with no human annotation,” IEEE Trans. Pattern Anal. Mach. Intell.,
vol. 41, no. 7, pp. 1655–1668, 2018.
[13] T. Weyand, I. Kostrikov, and J. Philbin, “Planet-photo geolocation with
convolutional neural networks,” in European Conf. on Computer Vision,
2016, pp. 37–55.
[14] B.-C. Chen and L. S. Davis, “Deep representation learning for metadata
verification,” in IEEE Winter Conf. on Applications of Computer Vision,
2019, pp. 73–82.
[15] S. Ginosar, K. Rakelly, S. Sachs, B. Yin, and A. A. Efros, “A century of
portraits: A visual historical record of american high school yearbooks,”
in IEEE Intl. Conf. on Computer Vision Workshops, 2015, pp. 1–7.
[16] T. Salem, S. Workman, M. Zhai, and N. Jacobs, “Analyzing human
appearance as a cue for dating images,” in IEEE Winter Conf. on
Applications of Computer Vision, 2016, pp. 1–8.
[17] Y. Jae Lee, A. A. Efros, and M. Hebert, “Style-aware mid-level repre-
sentation for discovering visual connections in space and time,” in IEEE
Intl. Conf. on Computer Vision, 2013, pp. 1857–1864.
[18] S. Vittayakorn, A. C. Berg, and T. L. Berg, “When was that made?”
in IEEE Winter Conf. on Applications of Computer Vision, 2017, pp.
[19] S. Lee, N. Maisonneuve, D. Crandall, A. Efros, and J. Sivic, “Linking
past to present: Discovering style in two centuries of architecture,” in
IEEE Intl. Conf. on Computational Photography, 2015.
[20] T.-H. Tsai, W.-C. Jhou, W.-H. Cheng, M.-C. Hu, I.-C. Shen, T. Lim, K.-
L. Hua, A. Ghoneim, M. A. Hossain, and S. C. Hidayati, “Photo sundial:
estimating the time of capture in consumer photos,” Neurocomputing,
vol. 177, pp. 529–542, 2016.
[21] B. Fernando, D. Muselet, R. Khan, and T. Tuytelaars, “Color features for
dating historical color images,” in IEEE Intl. Conf. on Image Processing,
2014, pp. 2589–2593.
[22] P. Martin, A. Doucet, and F. Jurie, “Dating color images with ordinal
classification,” in ACM Intl. Conf. on Multimedia Retrieval, 2014, p.
[23] F. Palermo, J. Hays, and A. Efros, “Dating historical color images,”
European Conf. on Computer Vision, pp. 499–512, 2012.
[24] A. Volokitin, R. Timofte, and L. Van Gool, “Deep features or not:
Temperature and time prediction in outdoor scenes,” in IEEE Conf. on
Computer Vision and Pattern Recognition Workshops, 2016, pp. 63–71.
[25] R. Baltenberger, M. Zhai, C. Greenwell, S. Workman, and N. Jacobs,
“A fast method for estimating transient scene attributes,” IEEE Winter
Conf. on Applications of Computer Vision, 2016.
[26] M. Zhai, T. Salem, C. Greenwell, S. Workman, R. Pless, and N. Jacobs,
“Learning geo-temporal image features,” in British Machine Vision
Conf., 2018.
[27] N. Jacobs, N. Roman, and R. Pless, “Consistent temporal variations in
many outdoor scenes,” in IEEE Conf. on Computer Vision and Pattern
Recognition, 2007, pp. 1–6.
[28] B. Thomee, D. A. Shamma, G. Friedland, B. Elizalde, K. Ni, D. Poland,
D. Borth, and L.-J. Li, “YFCC100M: The new data in multimedia
research,” Communications of the ACM, vol. 59, no. 2, pp. 64–73, 2016.
[29] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” in Intl. Conf. on Learning Representa-
tions, 2015.
[30] B. Zhou, A. Lapedriza, A. Khosla, A. Oliva, and A. Torralba, “Places: A
10 million image database for scene recognition,” IEEE Trans. Pattern
Anal. Mach. Intell., vol. 40, no. 6, pp. 1452–1464, 2017.
[31] K. He, X. Zhang, S. Ren, and J. Sun, “Identity mappings in deep residual
networks,” in European Conf. on Computer Vision, 2016, pp. 630–645.
[32] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet:
A large-scale hierarchical image database,” in IEEE Conf. on Computer
Vision and Pattern Recognition, 2009, pp. 248–255.
[33] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,
arXiv preprint arXiv:1412.6980, 2014.
[34] G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely
connected convolutional networks,” in IEEE Conf. on Computer Vision
and Pattern Recognition, 2017, pp. 4700–4708.
[35] M. D. Zeiler and R. Fergus, “Visualizing and understanding convolu-
tional networks,” in European Conf. on Computer Vision, 2014, pp. 818–
Rafael Padilha is currently pursuing his Ph.D. in Computer Science at the
Institute of Computing, University of Campinas, Brazil. Padilha received his
M.Sc. in Computing Science in 2017 from the same university. His research
interests include machine learning, computer vision, and digital forensics.
Tawfiq Salem is a visiting assistant professor in the Department of Computer
and Information Technology at Purdue University, USA. He received his Ph.D.
in Computer Science from University of Kentucky in 2019. His research
interests include machine learning, computer vision, and remote sensing.
Scott Workman received his Ph.D. in Computer Science from University
of Kentucky in 2018. His research interests include computer vision and
machine learning, specifically focused on understanding the connections
between images and their geo-temporal context.
Fernanda A. Andal´
ois a researcher associated with the Institute of Com-
puting, University of Campinas, Brazil. Andal´
o received a Ph.D. in Computer
Science from the same university in 2012, during which she was a visiting
researcher at Brown University. She is an IEEE member and was the 2016-
2017 Chair of the IEEE Women in Engineering (WIE) South Brazil Section.
Her research interests include machine learning and computer vision.
Anderson Rocha has been an associate professor at the Institute of Comput-
ing, the University of Campinas, Brazil, since 2009. Rocha received his Ph.D.
in Computer Science from the University of Campinas. His research interests
include machine learning, reasoning for complex data, and digital forensics.
He was the chair of the IEEE Information Forensics and Security Technical
Committee for 2019-2020 term. He is a Senior Member of IEEE.
Nathan Jacobs is a professor of Computer Science at the University of
Kentucky. He received his Ph.D. in Computer Science from Washington
University in St. Louis, MO. His research interests include learning-based
image understanding, remote sensing, and using crowd-sourced imagery for
Earth observation applications. He is a Senior Member of the IEEE.
... While their central focus was on mapping tasks, they demonstrated the ability to estimate a probability distribution over the capture time. More recent work, by Padilha et al. [123] has significantly improved the performance on both the problems of verifying whether a purported timestamp is correct and estimating the distribution over possibly valid timestamps. ...
Full-text available
As unconventional sources of geo-information, massive imagery and text messages from open platforms and social media form a temporally quasi-seamless, spatially multi-perspective stream, but with unknown and diverse quality. Due to its complementarity to remote sensing data, geo-information from these sources offers promising perspectives, but harvesting is not trivial due to its data characteristics. In this article, we address key aspects in the field, including data availability, analysis-ready data preparation and data management, geo-information extraction from social media text messages and images, and the fusion of social media and remote sensing data. We then showcase some exemplary geographic applications. In addition, we present the first extensive discussion of ethical considerations of social media data in the context of geo-information harvesting and geographic applications. With this effort, we wish to stimulate curiosity and lay the groundwork for researchers who intend to explore social media data for geo-applications. We encourage the community to join forces by sharing their code and data.
Full-text available
Image descriptors based on activations of Convolutional Neural Networks (CNNs) have become dominant in image retrieval due to their discriminative power, compactness of the representation, and the efficiency of search. Training of CNNs, either from scratch or fine-tuning, requires a large amount of annotated data, where high quality of the annotation is often crucial. In this work, we propose to fine-tune CNNs for image retrieval on a large collection of unordered images in a fully automatic manner. Reconstructed 3D models, obtained by the state-of-the-art retrieval and structure-from-motion methods, guide the selection of the training data. We show that both hard positive and hard negative examples, selected by exploiting the geometry and the camera positions available from the 3D models, enhance the performance in particular object retrieval. CNN descriptor whitening discriminatively learned from the same training data outperforms the commonly used PCA whitening. We propose a novel trainable Generalized-Mean (GeM) pooling layer that generalizes max and average pooling and show that it boosts retrieval performance. Applying the proposed method on VGG network achieves state-of-the-art performance on standard benchmarks: Oxford Buildings, Paris, and Holidays datasets.
Full-text available
Recent work has shown that convolutional networks can be substantially deeper, more accurate and efficient to train if they contain shorter connections between layers close to the input and those close to the output. In this paper we embrace this observation and introduce the Dense Convolutional Network (DenseNet), where each layer is directly connected to every other layer in a feed-forward fashion. Whereas traditional convolutional networks with L layers have L connections, one between each layer and its subsequent layer (treating the input as layer 0), our network has L(L+1)/2 direct connections. For each layer, the feature maps of all preceding layers are treated as separate inputs whereas its own feature maps are passed on as inputs to all subsequent layers. Our proposed connectivity pattern has several compelling advantages: it alleviates the vanishing gradient problem and strengthens feature propagation; despite the increase in connections, it encourages feature reuse and leads to a substantial reduction of parameters; its models tend to generalize surprisingly well. We evaluate our proposed architecture on five highly competitive object recognition benchmark tasks. The DenseNet obtains significant improvements over the state-of-the-art on all five of them (e.g., yielding 3.74% test error on CIFAR-10, 19.25% on CIFAR-100 and 1.59% on SVHN).
Advances in photo editing and manipulation tools have made it significantly easier to create fake imagery. Learning to detect such manipulations, however, remains a challenging problem due to the lack of sufficient amounts of manipulated training data. In this paper, we propose a learning algorithm for detecting visual image manipulations that is trained only using a large dataset of real photographs. The algorithm uses the automatically recorded photo EXIF metadata as supervisory signal for training a model to determine whether an image is self-consistent — that is, whether its content could have been produced by a single imaging pipeline. We apply this self-consistency model to the task of detecting and localizing image splices. The proposed method obtains state-of-the-art performance on several image forensics benchmarks, despite never seeing any manipulated images at training. That said, it is merely a step in the long quest for a truly general purpose visual forensics tool.
Deep residual networks have emerged as a family of extremely deep architectures showing compelling accuracy and nice convergence behaviors. In this paper, we analyze the propagation formulations behind the residual building blocks, which suggest that the forward and backward signals can be directly propagated from one block to any other block, when using identity mappings as the skip connections and after-addition activation. A series of ablation experiments support the importance of these identity mappings. This motivates us to propose a new residual unit, which further makes training easy and improves generalization. We report improved results using a 1001-layer ResNet on CIFAR-10/100, and a 200-layer ResNet on ImageNet.
Is it possible to build a system to determine the location where a photo was taken using just its pixels? In general, the problem seems exceptionally difficult: it is trivial to construct situations where no location can be inferred. Yet images often contain informative cues such as landmarks, weather patterns, vegetation, road markings, and architectural details, which in combination may allow one to determine an approximate location and occasionally an exact location. Websites such as GeoGuessr and View from your Window suggest that humans are relatively good at integrating these cues to geolocate images, especially en-masse. In computer vision, the photo geolocation problem is usually approached using image retrieval methods. In contrast, we pose the problem as one of classification by subdividing the surface of the earth into thousands of multi-scale geographic cells, and train a deep network using millions of geotagged images. While previous approaches only recognize landmarks or perform approximate matching using global image descriptors, our model is able to use and integrate multiple visible cues. We show that the resulting model, called PlaNet, outperforms previous approaches and even attains superhuman levels of accuracy in some cases. Moreover, we extend our model to photo albums by combining it with a long short-term memory (LSTM) architecture. By learning to exploit temporal coherence to geolocate uncertain photos, we demonstrate that this model achieves a 50% performance improvement over the single-image model.