Figure 2 - available via license: Creative Commons Zero 1.0
Content may be subject to copyright.
Applying location encoders to differentiate two visually similar species ((a)-(j)) or two visually similar land use types ((k)-(t)). Arctic fox and bat-eared fox might look very similar visually as shown in (a) and (f). However, they have different spatial distributions. (b) and (g) show their distinct patterns in species image locations. (c)-(e): The predicted distributions of Arctic fox from different location encoders (without images as input). (h)-(j): The predicted distributions of bat-eared fox. Similarly, it might be hard to differentiate factories/powerplants from multi-unit residential buildings only based on their overhead satellite imgeries as shown in (k) and (p). However, as shown in (l) and (q), they have very different global spatial distributions. (m)-(o) and (r)-(t) show the predicted spatial distributions of factories/powerplants and multi-unit residential buildings from different location encoders. We can see that while í µí±¤í µí±í µí±í µí± (Mac Aodha et al., 2019) produces a over-generalized spatial distribution, í µí± í µí±ℎí µí±í µí±í µí±í µí° ¶+ and
Source publication
Generating learning-friendly representations for points in space is a fundamental and long-standing problem in ML. Recently, multi-scale encoding schemes (such as Space2Vec and NeRF) were proposed to directly encode any point in 2D/3D Euclidean space as a high-dimensional vector, and has been successfully applied to various geospatial prediction an...
Contexts in source publication
Context 1
... and (r)-(t) show the predicted spatial distributions of factories/powerplants and multi-unit residential buildings from different location encoders. We can see that while í µí±¤í µí±í µí±í µí± (Mac Aodha et al., 2019) produces a over-generalized spatial distribution, í µí± í µí±ℎí µí±í µí±í µí±í µí° ¶+ and í µí±í µí± í µí± (our model) produces more compact and fine-grained distributions on the polar region and in data sparse areas such as Africa (See Figure 2g-2j). í µí±í µí±í µí±í µí± (Mai et al., 2020b) is between the two. ...
Context 2
... demonstrate the effectiveness of Sphere2Vec on geo-aware image classification tasks including fine-grained species recognition ( Chu et al., 2019;Mac Aodha et al., 2019;Mai et al., 2020b), Flickr image recognition ( Tang et al., 2015;Mac Aodha et al., 2019), and remote sensing image classification ( Christie et al., 2018;Ayush et al., 2020). Figure 2c-2e and 2h-2j show the predicted species distributions of Arctic fox and bat-eared fox from three different models. Figure 2m-2o and 2r-2t show the predicted land use distributions of factory or powerplant and multiunit residential building from three different models. ...
Context 3
... 2c-2e and 2h-2j show the predicted species distributions of Arctic fox and bat-eared fox from three different models. Figure 2m-2o and 2r-2t show the predicted land use distributions of factory or powerplant and multiunit residential building from three different models. In summary, the contributions of our work are: ...
Context 4
... results on three datasets are shown here including BirdSnap †, NABirds †, and iNat2018. For each model, we vary the depth of its í µí°í µí° í µí± í µí± í µí± (), i.e., ℎ = [1, 2, 3, 4]. The best evaluation results with each ℎ are reported. ...
Context 5
... the same practice of Figure 10, Figure 12 shows similar analysis results on the fMoW dataset. Figure 12a visualizes the sample locations in the fMoW validation dataset and Figure 12b shows the numbers of training and validation samples in each latitude band. ...
Context 6
... the same practice of Figure 10, Figure 12 shows similar analysis results on the fMoW dataset. Figure 12a visualizes the sample locations in the fMoW validation dataset and Figure 12b shows the numbers of training and validation samples in each latitude band. Similar to the iNat2017 dataset, we can see that for the fMoW dataset more samples are available in the North hemisphere, especially when í µí¼ > 20 • . ...
Context 7
... the same practice of Figure 10, Figure 12 shows similar analysis results on the fMoW dataset. Figure 12a visualizes the sample locations in the fMoW validation dataset and Figure 12b shows the numbers of training and validation samples in each latitude band. Similar to the iNat2017 dataset, we can see that for the fMoW dataset more samples are available in the North hemisphere, especially when í µí¼ > 20 • . ...
Context 8
... to Figure 10c, Figure 12c shows the Δí µí±í µí± í µí± = í µí±í µí± í µí± (í µí±í µí± í µí± ) − í µí±í µí± í µí± (í µí±í µí±í µí±í µí±) for each latitude-longitude cell. Red and blue color indicates positive and negative Δí µí±í µí± í µí± . ...
Context 9
... to Figure 10d, Figure 12d visualizes the Δí µí±í µí± í µí± between each model to í µí±í µí±í µí±í µí± in different latitude bands on the fMoW dataset. We can see that all Sphere2Vec models can outperform í µí±í µí±í µí±í µí± on all latitude bands. ...
Context 10
... µí±í µí± í µí± has a clear advantage over all the other models on all bands. Moreover, all Sphere2Vec models have clear advantages over í µí±í µí±í µí±í µí± near the North pole and South pole which further confirms our Hypothesis A. In latitude band í µí¼ ∈ [0 • , 10 • ) where we have fewer training samples (see Figure 12b), í µí±í µí± í µí± has clear advantages over other models which confirms our Hypothesis B. ...
Similar publications
It happens in the examination paper that text lines include inconsistent nonuniform word size, character erasure, diverse text length and dense long texts. This paper proposes an improved method for ViT to enhance its capability in recognizing text lines in handwritten Chinese examination papers. First, this method employs a segmentation method sui...
