The Study of the Video Encoder Efficiency in Decoder-side Depth Estimation Applications

The paper presents a study of a lossy compression impact on depth estimation and virtual view quality. Two scenarios were considered: the approach based on ISO/IEC 23090-12 coder-agnostic MPEG Immersive video standard, and the more general approach based on simulcast video coding. The commonly used compression techniques were tested: VVC (MPEG-I Part 3 / H.266), HEVC (MPEG H part 2 / H.265), AVC (MPEG 4 part 10 / H.264), MPEG-2 (MPEG 2 part 2 / H.262), AV1 (AOMedia Video 1), VP9 (AOMedia VP9). The quality of virtual views generated from the encoded stream was assessed by the IV-PSNR metric which is adapted to synthesized images. The results were presented as a relationship between virtual view quality and the quality of decoded real views. The main conclusion from performed experiments is that encoding quality and virtual view quality are encoder-dependent, therefore, the used video encoder should be carefully chosen to achieve the best quality in decoder-side depth estimation