Model-Aided Coding: Using 3-D Scene Models in Motion-Compensated Video Coding
ABSTRACT We show that traditional waveform-coding and 3-D modelbased coding are not competing alternatives but should be combined to support and complement each other. Both approaches are combined such that the generality of waveform coding and the e#ciency of 3-D model-based coding are available where needed. The combination is achieved by providing the block-based video coder with a second reference frame for prediction which is synthesized by the model-based coder. Since the coding gain of this approach is directly related to the quality of the synthetic frame, we have extended the model-aided coder  to exploit knowledge about illumination changes and multiple objects. Remaining model failures and objects that are not known at the decoder are handled by standard block-based motioncompensated prediction. A Lagrangian approach is employed to control the coder. Experimental results show that bit-rate savings of about 35 % are achieved at equal average PSNR when comparing the model-aided codec to TMN-10, the state-of-the-art test model of the H.263 standard.
- [Show abstract] [Hide abstract]
ABSTRACT: Known coding techniques for transmitting moving images at very low bit rates are explained by the source models on which these coding techniques are based. It is shown that with motion-compensated hybrid coding, object-based analysis-synthesis coding, knowledge-based coding and semantic coding, there is a consistent development of source models. In consequence these coding techniques can be combined in a layered coding system. From experimental results obtained for object-based analysis-synthesis, coding estimates for the coding efficiency of such a layered coding system are derived using head and shoulder video telephone test sequences. It is shown that an additional compression factor of about 3 can be expected with such a complex layered coding system, when compared to block-based hybrid coding.Signal Processing: Image Communication. 01/1995;
Conference Paper: Rate-distortion-efficient video compression using a 3-D head model[Show abstract] [Hide abstract]
ABSTRACT: In this paper we combine model-based video synthesis with block-based motion-compensated prediction (MCP). Two frames utilized for prediction where one frame is the previous decoded one and the other frame is provided by a model-based coder. The approach is integrated into an H.263-based video codec. Rate-distortion optimization is employed for the coding control. Hence, the coding efficiency does not decrease below H.263 even if the model based coder cannot describe the current scene. On the other hand, if the objects in the scene correspond to the model-based coder, significant gains in coding efficiency can be obtained compared to TMN-10, the test model of the H.263 standard. This is verified by experiments with natural head-and-shoulder sequences. Bit-rate savings of about 35% are achieved at equal average PSNR. When encoding at equal bit-rate, significant improvements in terms of subjective quality are visibleImage Processing, 1999. ICIP 99. Proceedings. 1999 International Conference on; 02/1999
- [Show abstract] [Hide abstract]
ABSTRACT: IMDSP Workshop '98, pp. 119-122, Alpbach, July 1998. In this paper we describe a model-based algorithm for the estimation of photometric properties in a scene recorded with a video camera. We focus on the coding of head-and-shoulder scenes at very low data-rates of about 1kbit/s. Facial animation parameters  spec-ifying facial expressions are estimated from video se-quences and transmitted to the decoder. There, the sequence is reconstructed by rendering a 3-D head model that is animated according to the facial pa-rameters. We show in this paper that the quality of the decoded images and the robustness of the motion estimation can be improved by considering photomet-ric effects. An illumination model based on Lambert reflection of directional colored light is added to the virtual scene and adapted to the current illumination condition. Experimental results show an improvement of about 1.4 dB in PSNR in comparison to simple am-bient illumination models.06/2002;
USING 3-D SCENE MODELS IN MOTION-COMPENSATED VIDEO CODING
Peter Eisert, Thomas Wiegand
University of Erlangen-Nuremberg
Information Systems Laboratory
ICIP’2000, Vancouver, September 2000.
We show that traditional waveform-coding and 3-D model-based
coding are not competing alternatives but should be combined to
support and complement each other. Both approaches are com-
bined such that the generality of waveform coding and the effi-
ciency of 3-D model-based coding are available where needed.
The combination is achieved by providing the block-based video
coder with a second reference frame for prediction which is syn-
thesized by the model-based coder. Since the coding gain of this
approach is directly related to the quality of the synthetic frame,
we have extended the model-aided coder  to cope with illumi-
nation changes and multiple objects. Remaining model failures
and objects that are not known at the decoder are handled by stan-
dard block-based motion-compensated prediction. Experimental
results show that bit-rate savings of up to 45 % are achieved at
equal average PSNR when comparing the model-aided codec to
TMN-10, the test model of the H.263 standard.
In recent years, several video coding standards such as H.261,
H.263 , MPEG-1, and MPEG-2 have been introduced, which
mainly address the compression of generic video data for digital
storage and communication services. These schemes are designed
on the basis of the statistics of the video signal without knowledge
of the semantic content and can therefore robustly be used for ar-
bitrary scenes. The design of model-based codecs  is based on
the semantics of the scene. Hence, if the semantic information
of the scene can be exploited, higher coding efficiency may be
achieved by model-based video codecs. Such a 3-D model-based
codec is restricted to scenes that can be composed of objects that
are known by the decoder. One typical class of scenes are head-
and-shoulder sequences which can be frequently found in applica-
tions such as video-telephone or video-conferencing systems. For
head-and-shoulder scenes, bit-rates of about 1 kbit/s with accept-
able quality can be achieved . This has also motivated the re-
cently determined Synthetic and Natural Hybrid Coding (SNHC)
part of the MPEG-4 standard .
The combination of traditional hybrid video coding methods
with model-based coding for higher coding efficiency has been
proposed by several researchers [6, 7]. In these approaches, the
mode decision is done for an entire frame and therefore the in-
formation from the 3-D model cannot be exploited if parts of the
frame cannot be described by the model-based coder.
In  we have presented an extension of an H.263 video coder
 that utilizes information from a model-based coder. Instead
of exclusively predicting the current frame of the video sequence
from the previous decoded frame, prediction from the synthetic
frame of the model-based coder is additionally allowed.
model-aided coder decides which prediction is efficient in terms
of rate-distortion performance. Hence, the coding efficiency does
not decrease below H.263 in the case the model-based coder can-
not describe the current scene.
In this paper, we extend the model-aided coder  to exploit
the efficiency of the model-based coder also for more sophisti-
cated video sequences with changing lighting conditions or mul-
tiple objects. Parameters describing the illumination in the scene
are estimated together with motion and deformation of the objects
resulting in more accurate model frames. Experimental results
demonstrate that the improved rate-distortion performance of the
model-aided codec can also be measured for head-and-shoulder
sequences with multiple objects and varying illumination.
2. VIDEO CODING ARCHITECTURE
Figure 1 shows the architecture of the proposed model-aided video
coder (MAC). This figure depicts the well-known hybrid video
Fig. 1. Structure of the model-aided video coder. Traditional
block-based MCP from the previous decoded frame is extended
by prediction from the current model frame.
coding loop that is extended by a model-based coder. The model-
based coder is running simultaneously to the hybrid coder, gener-
ating a synthetic model frame. This model frame is employed as a
second reference for block-based motion-compensated prediction
(MCP) in addition to the previous reconstructed reference frame.
For each macroblock, the video coder decides which of the two
frames to use for MCP. The bit-rate reduction for the proposed
scheme arises from those parts in the image that are well approxi-
mated by the model frame. For these blocks, the bit-rate required
for transmission of the motion vector and DCT-coefficients for the
residual coding is often highly reduced. For more details about the
rate-distortion optimized mode decision and the changes made to
the H.263+ syntax, see .
3. MODEL-BASED CODEC
the head model
about 1 kbit/s
Fig. 2. Basic structure of the model-based codec.
The structure of the model-based codec is depicted in Fig. 2.
The encoder analyzes the incoming frames and estimates the pa-
rameters of 3-D motion and deformation for all objects in the
scene. The deformations for the head model are represented by a
set of facial animation parameters (FAPs) according to the MPEG-
4standard . Motion anddeformation of other objects areparam-
eterized similarly. All parameters are entropy-encoded and trans-
mitted through the channel. The information from the 3-D models
and the facial expression synthesis are incorporated into the pa-
rameter estimation. The 3-D models describe the shape, texture,
and the motion constraints of the objects. For synthesis of facial
expressions, the transmitted FAPs are used to deform the 3-D head
model. The other objects are similarly moved and deformed in the
virtual scene. Finally, individual video frames are approximated
by simply rendering the 3-D scene.
In our model-based coder all parameters are estimated simul-
taneously usingahierarchical optical flowbased method. Inthe
optimization, an analysis-synthesis loop is employed. The mean
squared error between the rendered scene and the current video
frame is minimized by estimating changes of the FAPs and the pa-
rameters for the other objects. To simplify the optimization in the
high-dimensional parameter space, a linearized solution is directly
computed using information from the optical flow and motion con-
straints from the models. For more details about the parameter es-
timation and the generation of model frames, please refer to [4, 8].
3.1. Illumination Compensation
Themodel-aided coder presented in isnot capable of represent-
ing lighting changes correctly since the texture is not updated peri-
odically. Therefore, the coding gain is much smaller for video se-
quences with varying illumination. In order to exploit the informa-
tion from the model frame also for this class of sequences, we add
an illumination component to the scene model that describes the
photometric properties of object surfaces and light sources. This
way, the lighting in the model frame can be compensated towards
the original frame by changing the parameters of the photometric
The incident light in the original scene is assumed to consist
of ambient light and a directional light source with illumination di-
and thus the relation between the video frame intensity
?. The object surface is modeled by Lambertian reflection,
? and the
?????from the head model is
respectively  and the surface normal
head model. The Lambertian model is applied to all object pix-
els in the image. Each pixel contributes 3 equations for the 3
RGB color components with a common direction of the incident
light. 8 parameters (ambient light: 3, directional light: 3, illumi-
nation direction: 2) characterize the current illumination situation
for the entire object. By estimating these parameters with a linear
least-squares estimator as shown in , we are able to compensate
the dominant brightness differences of corresponding points in the
synthesized model frame and the camera frame. This improves the
quality of the model frame used for prediction in the model-aided
coder significantly if the illumination in the scene changes.
???control theintensityof ambient anddirectional light,
? is derived from the 3-D
3.2. Multiple Object Motion
So far, the model-based coder lacks the generality to cope with
multiple object motion and deformation. For example, the se-
quence Clapper Board (Fig. 4) shows a clap moving in front of
a person occluding most parts of the face. In order to exploit the
model frame also for multiple object sequences, some modifica-
tion to the parameter estimator are necessary. Two different cases
are distinguished: first, only the head and shoulder part is modeled
in the synthetic scene and, second, all objects are described by a
If no 3-D model exists for the additional objects, the model
frame does not show them and cannot be expected to improve the
prediction in the corresponding area. The rate-distortion decision
of the multi-frame predictor, however, ensures that the coding ef-
ficiency does not decrease below H.263 even for this case. On
the other hand, the model frame can still contribute to the predic-
tion of those parts in the image that are not occluded or uncov-
ered. This requires the motion estimator to determine the param-
eters also from partly occluded objects. The occluded parts are
detected using image gradients and intensity differences between
model and camera frame. They are classified as outliers in the
over-determined system of equations and not used for parameter
estimation. Additionally, only those FAPs are estimated that are
influenced by a sufficient number of equations. Otherwise they
remain constant until they are uncovered again.
Higher coding gains can be obtained if the additional objects
are also modeled in the synthetic scene. The parameter estimation
is performed in the same way for all objects and only the descrip-
tion for shape and motion/deformation constraints isadapted to the
individual object. For the clap in Fig. 4, e.g., a planar triangular
mesh is extracted from the first frame showing the entire object.
Five animation parameters are estimated: translation in three di-
rections, rotation in the image plane and opening of the clap. All
pixels showing a particular object contribute to the corresponding
system of linear equations for the estimation of animation param-
eters. The classification of the pixels to individual objects is deter-
mined exploiting knowledge from the synthetic 3-D scene.
4. EXPERIMENTAL RESULTS
Experiments are conducted with the two self-recorded natural CIF
sequences Clapper Board and Illumination. Rate-distortion curves
are measured by varying the DCT quantizer parameter over values
able producing the same PSNR values as at the encoder. The data
for the first intra-coded frame and the initial 3-D model are ex-
cluded from the results thus simulating steady-state behavior, i.e.,
we compare the inter-frame coding performance of both codecs
excluding the transition phase at the beginning of the sequence.
To specify the coding performance of the proposed model-aided
codec (MAC), we compare it to the H.263 test model, TMN-10
(Annexes D, F, I, J, and T enabled).
For the special case of head-and-shoulder sequences, bit-rate
savings of 35 % at the low bit-rate end corresponding to a coding
gain of 2-3 dB PSNR are reported [1, 8]. If the lighting in the
scene changes this coding gain is reduced, since the model frames
no longer represent the original video frames correctly. The addi-
tional estimation of the lighting situation, however, allows to adapt
the illumination condition in the synthetic scene to the real world.
The effectiveness of the illumination estimation is illustrated in
??. Bit-streams are generated that are decod-
Fig. 3. Rate-distortion plot for the sequence Illumination illustrat-
ing the achieved improvement when using an illumination estima-
Fig. 3 for the sequence Illumination. During the acquisition of this
sequence, one natural light source was moved to alter the illumina-
tion conditions. Two experiments are performed. For the first one,
only the FAPs are estimated to create a model frame. For the sec-
ond experiment, we additionally estimate the illumination param-
eters and generate motion- and illumination-compensated model
frames. As shown in Fig. 3, the gain in PSNR for the model-aided
coder compared to the TMN-10 is about 1 dB if no illumination
compensation is performed. An additional gain of about 1.5 dB is
achieved when exploiting illumination information.
In a second experiment, the influence of unknown objects in
the scene is investigated. Fig. 4 shows the first frames of the head-
and-shoulder sequence Clapper Board. During the first 50 frames,
the face is occluded by an object that cannot be represented by the
3-D models available at the decoder. As a result, the correspond-
Fig. 4. Frames 0, 11, 22, 33, 44, and 55 of the sequence Clapper
ing model frames do not contain this additional object as shown
in Fig. 7 a). Since prediction from the previous decoded frame
and residual coding provides uswithrobustnessagainst model fail-
ures, the model-aided coder represents the entire frame correctly
(Fig. 7 b)). The coding efficiency of the model-aided coder, how-
ever, drops down during the first frames as shown in the tempo-
ral evolution of the PSNR in Fig. 5. If the face is visible again
the model frame can be exploited and the PSNR recovers showing
high coding gains. The overall rate-distortion performance for the
050 100 150200250
MAC 2 models
Fig. 5. Temporal evolution of PSNR for the sequence Clapper
Board. Both coders use a DCT quantizer parameter of 31.
entire sequence is depicted in Fig. 6. Bit-rate savings of 33 % at
the low bit-rate end are achieved. The quality of the reconstructed
MAC 2 models
Fig. 6. Rate-distortion plot for sequence Clapper Board.
frames is illustrated in Fig. 7. Image b) shows frame 54 encoded
with the model-aided coder, while c) corresponds to the TMN-10
coder at the same bit-rate. Figure 7 d) shows a frame from the
TMN-10 coder that has the same PSNR as the model-aided frame.
Even though the PSNR is the same, the subjective quality of the
Fig. 7. Frame 54 of the sequence Clapper Board. a) Model frame
without clapper board; b) MAC, 36.1 dB, 2900 bits; c) TMN-10
with same average bit-rate as MAC, 32.7 dB, 3200 bits; d) TMN-
10 with same average PSNR as MAC, 36.0 dB, 5800 bits.
reconstructed frame from the model-aided coder is higher since
facial features are reproduced more accurately and with less arti-
facts. The difference is even more striking when viewing motion
In a third experiment, the clap in the sequence Clapper Board
is described by an additional 3-D model placed in the synthetic
scene. This model is manually acquired using the texture from
one frame that shows the entire clap. Motion and deformation
parameters are estimated for both objects using the approach in
Section 3. With these parameters, model frames are generated that
represent all objects in the scene. Running the model aided coder
withthesemodel framesresultsinamuchhigher PSNRforthefirst
framescompared tothecasewhen usingonly thehead model. This
is illustrated in the upper curve of Fig. 5. At the low bit-rate end,
an average gain of 4.6 dB is achieved for the first 60 frames when
using two models while the use of a single head model results in a
gain of 1.2 dB PSNR. The overall rate-distortion performance for
the entire sequence is depicted in Fig. 6. Bit-rate savings of 45 %
corresponding to a coding gain of about 3.5 dB are achieved.
The proposed model-aided codec which combines model-based
video coding with block-based motion-compensated prediction
yields a superior video coding scheme for head-and-shoulder se-
quences. The coding efficiency of this codec can be further in-
creased by improving the 3-D models describing the scene. We
have extended the model-aided codec to exploit the efficiency
of the model-based codec also for more sophisticated video se-
quences with changing lighting conditions or multiple objects. Pa-
rameters describing the illumination in the scene are estimated
together with motion and deformation of the objects resulting in
more accurate model frames. Experiments have shown that bit-
rate savings of up to 45 % can be achieved at equal average PSNR.
 P. Eisert, T. Wiegand, and B. Girod, “Rate-distortion-efficient video
compression using a 3-D head model,” in Proc. International Confer-
ence on Image Processing (ICIP), Kobe, Japan, Oct. 1999, vol. 4, pp.
 ITU-T Recommendation H.263 Version 2 (H.263+), “Video Coding
for Low Bitrate Communication,” Jan. 1998.
 D. E. Pearson, “Developments in model-based video coding,” Pro-
ceedings of the IEEE, vol. 83, no. 6, pp. 892–906, June 1995.
 P. Eisert and B. Girod, “Analyzing facial expressions for virtual con-
ferencing,” IEEE Computer Graphics and Applications, vol. 18, no.
5, pp. 70–78, Sep. 1998.
 ISO/IEC FDIS 14496-2, Generic Coding of audio-visual objects:
(MPEG-4 video), Final Draft International Standard, Document
 M. F. Chowdhury, A. F. Clark, A. C. Downton, E. Morimatsu, and
D. E. Pearson, “A switched model-based coder for video signals,”
IEEE Transactions on Circuits and Systems for Video Technology, vol.
4, no. 3, pp. 216–227, June 1994.
 H. G. Musmann, “A layered coding system for very low bit rate video
coding,” Signal Processing: Image Communication, vol. 7, no. 4-6,
pp. 267–278, Nov. 1995.
 P. Eisert, T. Wiegand, and B. Girod, “Model-aided coding: A new ap-
proach to incorporate facial animation into motion-compensated video
coding,” IEEE Transactions on Circuits and Systems for Video Tech-
nology, vol. 10, no. 3, pp. 344–358, Apr. 2000.
 P.Eisert and B. Girod, “Model-based coding of facial image sequences
at varying illumination conditions,” in Proc. 10th Image and Multidi-
mensional Digital Signal Processing Workshop IMDSP ’98, Alpbach,
Austria, Jul. 1998, pp. 119–122.