Contextual Siamese Learning Framework for Efficient Stereo Video Compression

With the rapid advancement of intelligent applications such as autonomous driving and virtual reality， stereo video has been widely used for environmental perception and immersive interaction because it simultaneously captures appearance information and depth cues. However， stereo video contains substantial redundancy in both the temporal and inter-view dimensions， and its data volume increases significantly with higher resolutions and frame rates， placing considerable demands on storage and transmission systems. Consequently， efficient stereo video compression has become an important research topic in video coding and multimedia signal processing. Compared with monocular video， stereo video involves not only more complex temporal dependencies but also significant inter-view redundancy， making redundancy modeling and efficient compression more challenging. Existing stereo video compression methods， including traditional multiview coding frameworks and end-to-end learned codecs， typically reduce redundancy through explicit residual coding and disparity prediction. However， at the probabilistic modeling level， these approaches often fail to fully exploit the statistical dependencies among multiple sources of contextual information. Their coding architectures generally rely on cascaded， multi-stage prediction and residual modules and lack an explicit and structured representation of the conditional dependencies between the left and right views， which limits model expressiveness. To address these limitations， this paper proposes a contextual Siamese learning framework for stereo video compression. The proposed dual-view coding architecture jointly models temporal and inter-view context in the feature space， while a conditional coding network adaptively estimates the probability distribution of the features to be encoded. This design effectively reduces temporal and inter-view redundancy without explicit residual computation. Experimental results on the Cityscapes and KITTI 2015 stereo video datasets demonstrate that the proposed method outperforms conventional stereo video codecs and representative learning-based approaches in rate-distortion performance， significantly improving the compression efficiency of end-to-end stereo video coding.