Video Object Detection From Compressed Formats for Modern Lightweight Consumer Electronics

The rapid rise of technological advancements led to the increased consumption of electronic gadgets. This change expedited the requirement for sustainable technologies to meet the growing consumer requirements with minimum computational costs. Nowadays, video content shares a large proportion of the Internet bandwidth. Object Detection from the videos is essential in various real-time applications. Traditionally, the videos are decoded to the raw format for detection tasks. This analytics process can be more efficient if the detection tasks are carried out from compressed video formats instead of raw video. The compressed format of the videos, produced by modern deep learning-based approaches, contains both semantic and motion information in easily consumable formats. Based on the same notion, a video compression cum object detection network has been proposed in this paper, which consumes the compressed videos for carrying out detection tasks. The proposed network comprises an already-designed video compression network, which has been extended to incorporate object detection capabilities. The proposed network has been experimented with a standard ImageNet VID dataset, and the results show fast and efficient object detection from the compressed videos. Coupled with temporal features, the proposed model achieves significantly better mAP of 44.3 w.r.t. 36.7 and 39.6 from YOLOv5-s and YOLOX-s models, respectively. The comparative results have shown incremental improvement in the detection tasks from the compressed videos, making it sustainable for its application in modern lightweight consumer electronic devices.