Research map / Video / Video-language understanding
Sparse-Dense Side-Tuner for efficient Video Temporal Grounding
David Pujol-Perich, Sergio Escalera, Albert Clapés
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 |
|---|---|---|---|
| reuse frozen AdaptivePool instead of CLS pooling for InternVideo2 features (ablation) | SDST · QVHighlights val | MR mAP avg: 50.53 -> 55.60 | Improved |
| reuse frozen AdaptivePool instead of average pooling (ablation) | SDST · QVHighlights val | MR mAP avg: 53.44 -> 55.60 | Improved |
| sample intermediate layers from shallower depth ([10,20,30,40] vs [37,38,39,40]) (ablation) | SDST · QVHighlights val | MR mAP avg: 55.60 -> 52.47 | Hurt |
| use intermediate-layer features instead of last-layer features for refinement steps (ablation) | SDST · QVHighlights val | MR mAP | Mixed |
| replace deformable cross-attention with RDSA (ablation) | SDST · QVHighlights val | MR R1@0.7 and mAP: 58.19 -> 60.90 R1@0.7; 54.27 -> 55.60 mAP | Improved |
| replace standard cross-attention with RDSA (ablation) | SDST · QVHighlights val | MR mAP avg: 42.72 -> 55.60 | Improved |
| add MixedInit decoder query initialization to deformable CA (ablation) | Def. CA · QVHighlights val | MR mAP avg: 54.27 -> 54.94 | No clear effect |
| add PureInit decoder query initialization to deformable CA (ablation) | Def. CA · QVHighlights val | MR mAP avg: 54.27 -> 52.92 | Hurt |
| condition sparse stream on post CA+SA+FFN output instead of raw video representation (ablation) | SDST · QVHighlights val | MR mAP avg: 50.74 -> 55.60 | Improved |
| add cross-attention (CA) module to sparse stream (ablation) | SDST · QVHighlights val | MR mAP avg: 54.47 -> 55.60 | Improved |
| add self-attention (SA) module to sparse stream (ablation) | SDST · QVHighlights val | MR mAP avg: 54.32 -> 55.60 | Improved |
| add RDSA module to sparse stream (ablation) | SDST · QVHighlights val | MR mAP avg: 29.09 -> 55.60 | 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.
- R2-Tuning — R^2-Tuning: Efficient Image-to-Video Transfer Learning for Video Temporal Grounding (ECCV 2024) · better on 5 of 5 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
- R^2-Tuning: Efficient Image-to-Video Transfer Learning for Video Temporal Grounding (ECCV 2024 · similarity 0.90)
- Empower Words: DualGround for Structured Phrase and Sentence-Level Temporal Grounding (NeurIPS 2025 · similarity 0.89)
- DeCafNet: Delegate and Conquer for Efficient Temporal Grounding in Long Videos (CVPR 2025 · similarity 0.89)
- SnAG: Scalable and Accurate Video Grounding (CVPR 2024 · similarity 0.88)
- Bridging the Gap: A Unified Video Comprehension Framework for Moment Retrieval and Highlight Detection (CVPR 2024 · similarity 0.87)
- Universal Video Temporal Grounding with Generative Multi-modal Large Language Models (NeurIPS 2025 · 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.
