Distributed Coding of Omnidirectional Images
Distributed source coding (DSC) refers to the independent encoding and joint encoding of correlated sources. DSC relies on the theoretical background for distributed compression proposed by Slepian-Wolf and Wyner-Ziv. Although the theoretical foundations have been defined in the late seventies, the first constructive design of a DSC scheme using channel codes has been formulated only recently. After that DSC gained a lot of attention among the researchers particularly in designing low complexity encoders for sensor networks.
We considered the scenario where the omnidirectional cameras are distributed in the 3D scene. In most practical scenarios, the omnidirectional images are correlated due to overlapping views. For example, the object present in the spherical images could be displaced. We propose DSC principles to exploit the correlation among the omnidirectional images. More details are available in the following papers.
- V. Thirumalai, I. Tosic and P. Frossard, Distributed coding of multiresolution omnidirectional images, Proc. IEEE Int. Conf. ICIP, pp. 345-348, Sept. 2007.
- V. Thirumalai, I. Tosic and P. Frossard, Balanced Distributed Coding of Omnidirectional Images, Proceedings of VCIP, VCIP 2008.
- V. Thirumalai, I. Tosic and P. Frossard, Symmetric Distributed coding of Stereo Omnidirectional Images, Special issue in Distributed coding, Signal Processing: Image Communication, Vol. 23, June 2008.
Distributed Representation from Compressed Linear Measurements
The distributed representation of correlated images is an important challenge in applications such as multi-view imaging in camera networks or low complexity video coding. We consider the problem of finding an efficient distributed representation and joint reconstruction solution for correlated images, where the common objects are displaced due to the viewpoint changes or motion in dynamic scenes. In particular, we are interested in finding an efficient distributed representation when the images are given under the form of few quantized linear measurements computed by very simple sensors.
In contrary to most distributed compressive schemes in the literature, we propose to estimate the correlation information prior to image reconstruction for improved robustness at low coding rates. We then use the estimated correlation information in a novel joint reconstruction algorithm based on a convex optimization framework that decodes the correlated images from the quantized measurements. More details are available in the following papers.
- Distributed Representation of Geometrically Correlated Images with Compressed Linear Measurements, accepted to IEEE Transactions on Image Processing, 2012.
- Image Reconstruction from Compressed Linear Measurements with Side Information, ICIP, 2011. .
- Correlation Estimation from Compressed Images, accepted to Journal of Visual Communication and Image Representation, 2012.
Joint reconstruction from Compressed Images
In state-of-the-art distributed coding schemes, usually a feedback channel is used in order to precisely control the Slepian-Wolf coding rate. Unfortunately, this results in high latency and bandwidth usage due to the multiple requests from the joint decoder, and therefore can hardly be used in real time applications. Also in most cases, a separate encoding rate control module is needed to precisely control the Slepian-Wolf coding rate which makes the encoding complexity non-negligible.
In our framework, we therefore omit the Slepian-Wolf encoder and we propose a rate balanced distributed coding scheme, where the compressed images (e.g., SPIHT-based compression) are directly transmitted to the joint decoder. The central decoder now builds a correlation model from the compressed images, which is eventually used to jointly decode a pair of images. We propose a novel joint reconstruction algorithm based on a convex optimization problem and we solve it effectively using proximal splitting methods.
Joint Reconstruction of Correlated Images from Compressed Images, submitted to European Signal Processing Conference (EUSIPCO), 2012..
Depth Estimation from Compressed Images
Dense correspondence problem or stereo matching is the one of active research areas in computer vision and finds an important step in 3D reconstruction, image based rendering, camera calibration etc. Several algorithms have been proposed in the literature to compute the depth map from the stereo images, but these algorithms compute the depth map from the original images by assuming that the vision sensor encodes the information without any loss. In other words, the depth is computed at the joint decoder, by neglecting the distortion due to compression. In practice the images are often compressed to save on transmission costs. In this work we propose a novel rate allocation scheme to compute the 3D structure of the scene from compressed stereo images, captured by the distributed vision sensor networks. More details are available in the following paper.
V. Thirumalai, and P. Frossard, Bit Rate Allocation for Disparity Estimation from Compressed Images, Proceedings of PCS 2009.