Yu-Wei Chao, Sudheendra Vijayanarasimhan, Bryan Seybold, David A. Ross, Jia Deng, Rahul Sukthankar
We propose TAL-Net, an improved approach to temporal action localization in video that is inspired by the Faster R-CNN object detection framework. TAL-Net addresses three key shortcomings of existing approaches: (1) we improve receptive field alignment using a multi-scale architecture that can accommodate extreme variation in action durations; (2) we better exploit the temporal context of actions for both proposal generation and action classification by appropriately extending receptive fields; and (3) we explicitly consider multi-stream feature fusion and demonstrate that fusing motion late is important. We achieve state-of-the-art performance for both action proposal and localization on THUMOS'14 detection benchmark and competitive performance on ActivityNet challenge.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Video | THUMOS’14 | Avg mAP (0.3:0.7) | 39.8 | TAL-Net |
| Video | THUMOS’14 | mAP IOU@0.1 | 59.8 | TAL-Net |
| Video | THUMOS’14 | mAP IOU@0.2 | 57.1 | TAL-Net |
| Video | THUMOS’14 | mAP IOU@0.3 | 53.2 | TAL-Net |
| Video | THUMOS’14 | mAP IOU@0.4 | 48.5 | TAL-Net |
| Video | THUMOS’14 | mAP IOU@0.5 | 42.8 | TAL-Net |
| Video | THUMOS’14 | mAP IOU@0.6 | 33.8 | TAL-Net |
| Video | THUMOS’14 | mAP IOU@0.7 | 20.8 | TAL-Net |
| Temporal Action Localization | THUMOS’14 | Avg mAP (0.3:0.7) | 39.8 | TAL-Net |
| Temporal Action Localization | THUMOS’14 | mAP IOU@0.1 | 59.8 | TAL-Net |
| Temporal Action Localization | THUMOS’14 | mAP IOU@0.2 | 57.1 | TAL-Net |
| Temporal Action Localization | THUMOS’14 | mAP IOU@0.3 | 53.2 | TAL-Net |
| Temporal Action Localization | THUMOS’14 | mAP IOU@0.4 | 48.5 | TAL-Net |
| Temporal Action Localization | THUMOS’14 | mAP IOU@0.5 | 42.8 | TAL-Net |
| Temporal Action Localization | THUMOS’14 | mAP IOU@0.6 | 33.8 | TAL-Net |
| Temporal Action Localization | THUMOS’14 | mAP IOU@0.7 | 20.8 | TAL-Net |
| Zero-Shot Learning | THUMOS’14 | Avg mAP (0.3:0.7) | 39.8 | TAL-Net |
| Zero-Shot Learning | THUMOS’14 | mAP IOU@0.1 | 59.8 | TAL-Net |
| Zero-Shot Learning | THUMOS’14 | mAP IOU@0.2 | 57.1 | TAL-Net |
| Zero-Shot Learning | THUMOS’14 | mAP IOU@0.3 | 53.2 | TAL-Net |
| Zero-Shot Learning | THUMOS’14 | mAP IOU@0.4 | 48.5 | TAL-Net |
| Zero-Shot Learning | THUMOS’14 | mAP IOU@0.5 | 42.8 | TAL-Net |
| Zero-Shot Learning | THUMOS’14 | mAP IOU@0.6 | 33.8 | TAL-Net |
| Zero-Shot Learning | THUMOS’14 | mAP IOU@0.7 | 20.8 | TAL-Net |
| Action Localization | THUMOS’14 | Avg mAP (0.3:0.7) | 39.8 | TAL-Net |
| Action Localization | THUMOS’14 | mAP IOU@0.1 | 59.8 | TAL-Net |
| Action Localization | THUMOS’14 | mAP IOU@0.2 | 57.1 | TAL-Net |
| Action Localization | THUMOS’14 | mAP IOU@0.3 | 53.2 | TAL-Net |
| Action Localization | THUMOS’14 | mAP IOU@0.4 | 48.5 | TAL-Net |
| Action Localization | THUMOS’14 | mAP IOU@0.5 | 42.8 | TAL-Net |
| Action Localization | THUMOS’14 | mAP IOU@0.6 | 33.8 | TAL-Net |
| Action Localization | THUMOS’14 | mAP IOU@0.7 | 20.8 | TAL-Net |