UnLoc: A Unified Framework for Video Localization Tasks

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}.

PDF Abstract ICCV 2023 PDF ICCV 2023 Abstract

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Zero-Shot Action Detection ActivityNet-1.3 UnLoc-B (75% seen split) mAP IOU@0.5 40.2 # 3
Zero-Shot Action Detection ActivityNet-1.3 UnLoc-B (50% seen split) mAP IOU@0.5 36.9 # 6
Zero-Shot Action Detection ActivityNet-1.3 UnLoc-L (50% seen split) mAP IOU@0.5 43.2 # 2
Zero-Shot Action Detection ActivityNet-1.3 UnLoc-L (75% seen split) mAP IOU@0.5 47.4 # 1
Temporal Action Localization ActivityNet-1.3 UnLoc-L mAP IOU@0.5 59.3 # 3
Natural Language Moment Retrieval ActivityNet Captions UnLoc-B R@1,IoU=0.5 48.0 # 4
R@1,IoU=0.7 29.7 # 4
R@5,IoU=0.5 81.5 # 1
R@5,IoU=0.7 61.4 # 2
Natural Language Moment Retrieval ActivityNet Captions UnLoc-L R@1,IoU=0.5 48.3 # 3
R@1,IoU=0.7 30.2 # 2
R@5,IoU=0.5 79.2 # 2
R@5,IoU=0.7 61.3 # 3
Moment Retrieval Charades-STA UnLoc-L R@1 IoU=0.5 60.8 # 1
R@1 IoU=0.7 38.4 # 2
R@5 IoU=0.5 88.2 # 3
R@5 IoU=0.7 61.1 # 1
Moment Retrieval Charades-STA UnLoc-B R@1 IoU=0.5 58.1 # 6
R@1 IoU=0.7 35.4 # 6
R@5 IoU=0.5 87.4 # 4
R@5 IoU=0.7 59.1 # 2
Action Segmentation COIN UnLoc-L Frame accuracy 72.8 # 1
Moment Retrieval QVHighlights UnLoc-L R@1 IoU=0.5 66.1 # 2
R@1 IoU=0.7 46.7 # 10
Moment Retrieval QVHighlights UnLoc-B R@1 IoU=0.5 64.5 # 6
R@1 IoU=0.7 48.8 # 3

Methods