We propose an action detection system for detecting human and vehicle actions in long untrimmed videos, submitted for the TRECVID Activities in Extended Video (ActEV) 2020 challenge. It utilizes an object detection and tracking stage to divide the initial video into object tracks for all possible actors, followed by action localization to temporally localize and classify all actions within these tracks. Finally, we conduct several experiments into spatial and
temporal relation modeling, both showing limited performance improvement, but demonstrating the possibility of similar approaches for future video action detection research.
Besides the VIRAT dataset utilized for the challenge, we utilize networks pretrained on the ImageNet and ActivityNet datasets. Summaries of the different submitted runs are as follows:
• 22342 - TTA-baseline: Standard two-stage system without any relation modeling
• 22442 - TTA-SRM: Same as baseline, but utilizing spatial relation modeling post-processing
• 22658 - TTA-SF2: System using multiple sampling rates for temporal action localization
• 22657 - TTA-SF: Same as SF2, but utilizing spatial relation modeling
From the run results, we can see that utilizing the multi-sampling rate action localization slightly improves performance, while the relation modeling decreases performance, contrary to our validation experiments. This seems to indicate that our relation modeling is still premature.
各種検索
サポート
ヘルプ