Object Based Coding of Video Sequences at Low Bit Rates Using Adaptively Trained Neural Networks

Unsupervised video object segmentation is proposed in this paper, using an adaptively trained neural network structure followed by a face and body detection scheme. The latter uses probabilistic modeling for applying the face and body detection task. The algorithm is incorporated along with a rate control mechanism, which allocates more bits to regions of importance, such as humans in video conferencing applications than to unimportant ones. The compatibility to block-based encoders, such as the MPEG-1/2 and the H.263, is retained, so that the proposed scheme can further improve their coding efficiency.