Investigation on Error-Tolerability Enhancement of Videos via Re-Encoding for Computer Vision: A Case Study on Object Detection
Tong-Yu Hsieh, Jun-Tsung Wu · 2021
Run-time video re-encoding via adjusting the Group of Pictures (GOP) size has been shown to be able to effectively enhance the error-tolerability of videos. However, this study was done by focusing on the human visual system’s perception to video errors. In this paper we investigate on this issue by taking machine’s tolerability of errors into consideration. Objection detection carried out by yolov3 is employed as a case study. In this study, errors are injected into a benchmark video to generate more than 100,000 erroneous videos, and then yolov3 is employed to evaluate the quality (acceptability) of these videos. The results demonstrate that video re-encoding can have higher application efficiency (e.g., larger GOP size) for computer vision, mainly because that the error-tolerability of machines is larger than that of humans. This can lead to better video encoding efficiency (smaller video size) and lower implementation cost for re-encoding.