Research map / Vision-language and multimodal / CLIP and zero-shot
Zero-Shot Temporal Action Detection via Vision-Language Prompting
Sauradip Nag, Xiatian Zhu, Yi-Zhe Song, Tao Xiang
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 |
|---|---|---|---|
| add Maskformer decoder representation masking (ablation) | STALE · ActivityNet v1.3 zero-shot, 75% seen | avg mAP: 10.9 -> 24.9 | Improved |
| use Maskformer decoder instead of 1-D CNN for masking (ablation) | STALE · ActivityNet v1.3 zero-shot, 75% seen | avg mAP: 20.1 -> 24.9 | Improved |
| increase decoder queries from 5 to 100 (ablation) | STALE · ActivityNet v1.3 zero-shot, 75% seen | avg mAP: 24.4 -> 25.1 | Improved |
| add text encoder (ablation) | STALE · ActivityNet v1.3 zero-shot, 75% seen | avg mAP: 11.5 -> 24.9 | Improved |
| fine-tune the CLIP text encoder (ablation) | STALE · ActivityNet v1.3 zero-shot, 75% seen | avg mAP: 22.8 -> 24.9 | Improved |
| use Transformer instead of MS-TCN for temporal modeling (ablation) | STALE · ActivityNet v1.3 zero-shot, 75% seen | avg mAP: 21.4 -> 24.9 | Improved |
| use Transformer instead of 1D CNN (ablation) | STALE · ActivityNet v1.3 zero-shot, 75% seen | avg mAP: 19.3 -> 24.9 | Improved |
| add vision-to-language cross-modal adaptation (ablation) | STALE · ActivityNet v1.3 zero-shot, 75% seen | avg mAP: 23.5 -> 24.9 | Improved |
| increase context tokens from 10 to 50 (ablation) | STALE · ActivityNet v1.3 zero-shot, 75% seen | avg mAP: 24.2 -> 24.9 | Improved |
| increase context tokens from 50 to 70 (ablation) | STALE · ActivityNet v1.3 zero-shot, 75% seen | avg mAP: 24.9 -> 24.8 | No clear effect |
| use parallel mask/classification (STALE) instead of proposal-based B-I (ablation) | STALE · ActivityNet v1.3 zero-shot, 75% seen | avg mAP drop from GT to predicted: 47.1 -> 19.9 for B-I vs 42.3 -> 24.9 for STALE | Improved |
| use STALE instead of EffPrompt | STALE · THUMOS14 zero-shot, 50% seen | avg mAP: 21.9 -> 22.2 | Improved |
Closest papers by meaning
- Test-Time Zero-Shot Temporal Action Localization (CVPR 2024 · similarity 0.92)
- Memory Matters: Boosting Training-Free Zero-Shot Temporal Action Localization with a Learnable Lookup Table (CVPR 2026 · similarity 0.91)
- Text-Infused Attention and Foreground-Aware Modeling for Zero-Shot Temporal Action Detection (NeurIPS 2024 · similarity 0.90)
- Semi-Supervised Temporal Action Detection with Proposal-Free Masking (ECCV 2022 · similarity 0.90)
- Dual DETRs for Multi-Label Temporal Action Detection (CVPR 2024 · similarity 0.88)
- Does Video-Text Pretraining Help Open-Vocabulary Online Action Detection? (NeurIPS 2024 · similarity 0.87)
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.
