Storage and retrieval of compressed video
Edward Te Chang, Avideh Zakhor · 1996
This work consists of three main parts. In the first part, we consider various aspects of video compression. We exploit motion in real video sequences within the context of subband coding by using two-dimensional nonseparable checker filters and adaptive one-dimensional two-tap temporal filters. Since the latter technique is shown to be more feasible for actual video encoding, we combine it with a multi-rate quantization strategy to form a scalable video subband encoder. Finally, we refine and optimize one aspect of our multi-rate quantization strategy known as significance map coding. The optimized map coding is integrated into a highly scalable subband codec and shown to increase the overall encoding and decoding speed by up to a factor of two, at a cost of one to two dB of decoded video quality. The second part of this work explores the issue of Variable Bit Rate (VBR) video storage. The fundamental issue in VBR data placement is the choice of blocks that are of constant playback duration or constant size. We call the first option constant time length (CTL) and the second constant data length (CDL); in addition, we consider a third hybrid option that stores CDL blocks but retrieves a variable number of blocks per user in each service round. By considering total disk and buffer cost to be our final quality metric, we show CTL and hybrid schemes to be cost-effective, but CDL is too expensive for long video sequences. We also address the issue of VBR admission control policies by comparing statistical and deterministic techniques. Statistical admission control uses statistics of the stored data to ensure that the probability of disk overload does not exceed a prespecified threshold. Deterministic control uses the actual stored video bit traces to regulate the number of admitted users. We show that one form of deterministic admission control, data-limit, is feasible and results in a reduction in cost and probability of overload. The final part addresses the issue of scalable video storage. To this end, we present the principle of constant frame grouping, placing an equal number of frames in the same rate layer together in each read unit. In comparison to other strategies, we find that constant frame grouping allows the video server to service more simultaneous users at a variety of user distributions. We then apply constant frame grouping to VBR video data storage using the CTL and hybrid data placement techniques. This allows the video server to provide graceful degradation in times of disk overload and also reduces the delay arising from the use of VCR functions in data-limit admission control. We extend our strategies to multiple disks and show that our interleaving technique serves more users and has lower delay than a comparable striped system.