Research map / Video / Action recognition
UniMD: Towards Unifying Moment Retrieval and Temporal Action Detection
Yingsen Zeng, Yujie Zhong, Chengjian Feng, Lin Ma
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| feed the query into the regression head (ablation) | UniMD · Ego4D val | TAD mAP: 21.86 -> 22.61 | Improved |
| use 3 fc layers instead of 1 in query transformation MLP (ablation) | UniMD · Ego4D val | MR R1@30: 12.65 -> 13.99 | Improved |
| pretrain on MR then finetune on TAD (ablation) | UniMD · Ego4D val | TAD mAP: 22.61 -> 22.69 | No clear effect |
| pretrain on TAD then finetune on MR (ablation) | UniMD · Ego4D val | MR R1@30: 13.99 -> 12.70 | Hurt |
| co-train TAD and MR with synchronized task sampling (ablation) | UniMD · Ego4D val | TAD mAP: 22.61 -> 24.12 | Improved |
| co-train with alternating task sampling instead of synchronized (ablation) | UniMD · Ego4D val | TAD mAP / MR R1@50: TAD mAP 22.61 -> 21.73; MR R1@50 9.34 -> 10.56 | Mixed |
| use loss weight 1-1 instead of 3-1 for co-training (ablation) | UniMD co-training · Ego4D val | TAD mAP: 24.12 -> 22.55 | Hurt |
| weight the TAD loss more heavily (3-1 vs 1-3) (ablation) | UniMD co-training · Ego4D val | TAD mAP: 19.61 -> 24.12 | Improved |
| co-train on only half of the TAD and MR training videos (ablation) | UniMD · ActivityNet / ActivityNet-Caption | TAD mAP: 38.60 -> 38.62 | Improved |
| co-train on only a quarter of the training videos (ablation) | UniMD · ActivityNet-Caption | MR R5@50: 77.28 -> 79.72 | Improved |
| increase co-training data volume from 0.25 to full (ablation) | UniMD · ActivityNet | TAD mAP: 36.78 -> 39.82 | Improved |
| add CLIP features to I3D features | UniMD · Charades TAD | mAP: 24.06 -> 26.53 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ActionFormer — ActionFormer: Localizing Moments of Actions with Transformers (ECCV 2022) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Dual DETRs for Multi-Label Temporal Action Detection (CVPR 2024 · similarity 0.88)
- An Empirical Study of End-to-End Temporal Action Detection (CVPR 2022 · similarity 0.88)
- PointTAD: Multi-Label Temporal Action Detection with Learnable Query Points (NeurIPS 2022 · similarity 0.88)
- Unsupervised Action Segmentation via Fast Learning of Semantically Consistent Actoms (AAAI 2024 · similarity 0.86)
- HyrCap: Hybrid Rank-Calibration of Action Proposals for Temporal Event Understanding (NeurIPS 2026 · similarity 0.86)
- MomentSeeker: A Task-Oriented Benchmark For Long-Video Moment Retrieval (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
