IN-Loop Filter for Object Mask Coding in Versatile Video Coding

Abstract

Translate Article

This paper explores the challenges and solutions in compressing object mask video, which can provide additional scene context for machine learning applications. Object masks, identifying specific objects or regions in images, are crucial for precise visual analysis. However, compressing them with video codecs tuned for human consumption may lead to detrimental artefacts for machine tasks. To alleviate such artefacts, a modification to the luma mapping and chroma scaling (LMCS) in-loop filter in the Versatile Video Coding standard is proposed, targeting the reconstruction of object mask video. By performing object mask reconstruction within the encoding loop, the encoder has access to correctly reference pictures for prediction. Consequently, coding performance is significantly enhanced. Objective evaluation against post-state-of-the-art processing mask reconstruction confirms a 42.0dB and 66.1dB improvement in Y-PSNR for random access and all intra encoding, respectively, while maintaining coding complexity similar to unmodified encoding without post-processing.