SPS, Wed 15/12/2010, TVNUM, 10:30 am:
Speaker: Amit Kumar K.C.
Title: "Optimal Dense Disparity Map Quantization and Residual Prediction for Lossless Stereo Image Coding".
This presentation is based on my thesis work, which is compression of stereo images.
Reproducing a natural and real scene as we see in the real world everyday is becoming more and more popular. Stereoscopic and multi-view techniques are used for this end. However, the fact that more information is displayed requires supporting technologies, such as digital compression, to ensure the storage and transmission of these sequences. Stereoscopic imaging has a wide range of applications such as digital cinema, medical imaging, satellite imagery, video games etc. All these applications require large amounts of storage. Consequently compression has become necessary.
Since the two images in stereo pair are slightly different, independent coding of both images is redundant. Besides, estimating one view from another enables high efficiency in compression of stereo pairs. Most of the traditional systems compute disparity between the reference image and the target image, and then residual image is computed using the target image and disparity compensated reference image. Thus, the bitrate of target image is distributed between disparity field and the residual image.
The principal aim of this work is to improve the exploitation of high redundancy in the stereo pair in lossless mode. A special attention has been paid so as to reduce overhead of side information. The work investigates a simple yet effective illumination compensation, based on homomorphic filtering. Besides, it employs existing optical flow method to compute the dense disparity map and then an optimal quantization is performed on a rate-distortion framework so as to minimize total required bitrate. Moreover, instead of just computing the difference for the residue, we compute the transformed residue by applying a filter to the disparity compensated image. The filter is so designed that it minimizes the variance of difference image. As the variance and entropy of residual image are monotonically related, the bitrate is reduced significantly at the cost of negligible overhead. Apart from these, a hybrid lifting scheme has also been explored which switches between a long filter (5/3 filter bank) for smooth regions and a short filter (Haar filter bank) for edges in transformed residual image.
Last updated September 25, 2013, at 09:16 AM