Research map / Video / Video-language understanding
Negative Sample Matters: A Renaissance of Metric Learning for Temporal Grounding
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 sentence negatives to mutual matching (ablation) | MMN · Charades-STA and ActivityNet-Captions | Rank n@m | Helped |
| Use inter-video negatives instead of only intra-video negatives (ablation) | MMN mutual matching · Charades-STA, especially | Rank n@m | Helped |
| Use both intra-video and inter-video negatives (ablation) | MMN mutual matching · Charades-STA and ActivityNet-Captions | Rank n@m | Helped |
| Use mutual matching without the original BCE loss (ablation) | MMN · Charades-STA and ActivityNet-Captions | Rank n@m | No clear effect |
| Sample the same number of negatives from both intra-video and inter-video candidates, rather than using only intra-video negatives (ablation) | MMN mutual matching · Charades-STA | R@1 | No clear effect |
| Sample the same number of negatives from both intra-video and inter-video candidates, rather than using only intra-video negatives (ablation) | MMN mutual matching · Charades-STA | R@5 | Helped |
| Use average pooling over sentence tokens instead of the [CLS] embedding (ablation) | MMN · temporal grounding benchmarks | grounding performance | Helped |
| Vary pair-discrimination margin, temperature, or loss weight (ablation) | MMN · Charades-STA | gap to baseline | No clear effect |
| Fuse IoU-regression and mutual-matching scores with a product (ablation) | MMN · temporal grounding experiments | grounding performance: about 0.5% | Helped |
| Replace ordered video clips with randomized visual input (ablation) | MMN · Charades-STA | R@1 (IoU=0.5): 37.5 vs 37.8 | Hurt |
| Use MMN's shared video features and late-fusion design instead of the 2D-TAN training setup | ActivityNet-Captions · ActivityNet-Captions | training time (GPU hours): 36 GPU hours -> 10 GPU hours | Helped |
Closest papers by meaning
- Multi-Pair Temporal Sentence Grounding via Multi-Thread Knowledge Transfer Network (AAAI 2025 · similarity 0.89)
- Empower Words: DualGround for Structured Phrase and Sentence-Level Temporal Grounding (NeurIPS 2025 · similarity 0.89)
- Universal Video Temporal Grounding with Generative Multi-modal Large Language Models (NeurIPS 2025 · similarity 0.88)
- Learning Multi-Scale Video-Text Correspondence for Weakly Supervised Temporal Article Gronding (AAAI 2024 · similarity 0.88)
- Multi-Scale Contrastive Learning for Video Temporal Grounding (AAAI 2025 · similarity 0.88)
- SnAG: Scalable and Accurate Video Grounding (CVPR 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.