Seam carving for semantic video coding

Compression standards such as H.264/AVC encode video sequences to maximize fidelity at a given bitrate. However, semantic-oriented and content-aware compression remains a challenge. In this paper, we propose a semantic video compression method using seam carving. Seam carving changes the dimension of an image/video with a non-uniform resampling of each row and column while keeping the rectangular shape of the image. Our main contribution is a new approach to identify areas where seams are concentrated. On the one hand, it allows to transmit supplemental seams data at low cost. On the other hand, seams can be synthesized at the decoder in order to recover the original frame size and to preserve the scene geometry. Experiments show that our seam carving method combined with standard H.264/AVC coding results in significant bitrate savings compared with the original H.264/AVC. Reported gains reach 39% at very high bitrates and 22% at very low bitrates. Furthermore, the reconstructed video has the same quality in semantically significant regions.