Pre-analysis-based adaptive temporal resampling for machine vision

Machine vision-oriented video compression is a fundamental infrastructure in the era of artificial intelligence. To improve the coding efficiency for various machine vision tasks, temporal redundancy is investigated and eliminated from the perspective of machine vision. Driven by the diversity of video content, a pre-analysisbased adaptive temporal resampling algorithm is proposed in which the resampling ratio is adaptively decided based on the machine-oriented temporal coherence (MTC) of the video. The proposed method, evaluated under the common test conditions (CTC) of video coding for machines defined by Joint Video Experts Team (JVET), significantly improves the coding performance for machine vision. For object detection task on SFU-HW dataset, the proposed algorithm achieves 67.65%, 3.13% and 18.45% and BD-rate savings under all intra, random access and low delay configurations, respectively. And for object tracking task on TVD dataset, 72.72%, 5.64% and 31.25% BD-rate savings are achieved under the same respective configurations.