Shen Yan, Xuehan Xiong, Arsha Nagrani, Anurag Arnab, Zhonghao Wang, Weina Ge, David Ross, Cordelia Schmid
While large-scale image-text pretrained models such as CLIP have been used for multiple video-level tasks on trimmed videos, their use for temporal localization in untrimmed videos is still a relatively unexplored task. We design a new approach for this called UnLoc, which uses pretrained image and text towers, and feeds tokens to a video-text fusion model. The output of the fusion module are then used to construct a feature pyramid in which each level connects to a head to predict a per-frame relevancy score and start/end time displacements. Unlike previous works, our architecture enables Moment Retrieval, Temporal Localization, and Action Segmentation with a single stage model, without the need for action proposals, motion based pretrained features or representation masking. Unlike specialized models, we achieve state of the art results on all three different localization tasks with a unified approach. Code will be available at: \url{https://github.com/google-research/scenic}.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Video | ActivityNet-1.3 | mAP IOU@0.5 | 59.3 | UnLoc-L |
| Video | ActivityNet Captions | R@1,IoU=0.5 | 48.3 | UnLoc-L |
| Video | ActivityNet Captions | R@1,IoU=0.7 | 30.2 | UnLoc-L |
| Video | ActivityNet Captions | R@5,IoU=0.5 | 79.2 | UnLoc-L |
| Video | ActivityNet Captions | R@5,IoU=0.7 | 61.3 | UnLoc-L |
| Video | ActivityNet Captions | R@1,IoU=0.5 | 48 | UnLoc-B |
| Video | ActivityNet Captions | R@1,IoU=0.7 | 29.7 | UnLoc-B |
| Video | ActivityNet Captions | R@5,IoU=0.5 | 81.5 | UnLoc-B |
| Video | ActivityNet Captions | R@5,IoU=0.7 | 61.4 | UnLoc-B |
| Temporal Action Localization | ActivityNet-1.3 | mAP IOU@0.5 | 59.3 | UnLoc-L |
| Zero-Shot Learning | ActivityNet-1.3 | mAP IOU@0.5 | 59.3 | UnLoc-L |
| Action Localization | ActivityNet-1.3 | mAP IOU@0.5 | 59.3 | UnLoc-L |
| Action Localization | COIN | Frame accuracy | 72.8 | UnLoc-L |
| Action Segmentation | COIN | Frame accuracy | 72.8 | UnLoc-L |
| Moment Retrieval | Charades-STA | R@1 IoU=0.5 | 60.8 | UnLoc-L |
| Moment Retrieval | Charades-STA | R@1 IoU=0.7 | 38.4 | UnLoc-L |
| Moment Retrieval | Charades-STA | R@5 IoU=0.5 | 88.2 | UnLoc-L |
| Moment Retrieval | Charades-STA | R@5 IoU=0.7 | 61.1 | UnLoc-L |
| Moment Retrieval | Charades-STA | R@1 IoU=0.5 | 58.1 | UnLoc-B |
| Moment Retrieval | Charades-STA | R@1 IoU=0.7 | 35.4 | UnLoc-B |
| Moment Retrieval | Charades-STA | R@5 IoU=0.5 | 87.4 | UnLoc-B |
| Moment Retrieval | Charades-STA | R@5 IoU=0.7 | 59.1 | UnLoc-B |
| Moment Retrieval | QVHighlights | R@1 IoU=0.5 | 66.1 | UnLoc-L |
| Moment Retrieval | QVHighlights | R@1 IoU=0.7 | 46.7 | UnLoc-L |
| Moment Retrieval | QVHighlights | R@1 IoU=0.5 | 64.5 | UnLoc-B |
| Moment Retrieval | QVHighlights | R@1 IoU=0.7 | 48.8 | UnLoc-B |