Research map / Video / Video-language understanding
ResidualViT for Efficient Temporally Dense Video Encoding
Mattia Soldan, Fabian Caba Heilbron, Bernard Ghanem, Josef Sivic, Bryan Russell
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 interleaving of full-encoded I-frames with token-reduced P-frames (N=2) (ablation) | token-reduced CLIP ViT-L/14 · Charades-STA NLTVG ablation (ViT-L/14) | R@1 IoU=0.5: 28.5 -> 38.9 | Improved |
| add interleaving (N=2) (ablation) | token-reduced CLIP ViT-L/14 · Charades-STA NLTVG ablation (ViT-L/14) | R@1 IoU=0.7: 14.5 -> 22.8 | Improved |
| add distilled residual tokenizer (ablation) | ResidualViT (reduction + interleave) · Charades-STA NLTVG ablation (ViT-L/14) | R@1 IoU=0.5: 38.9 -> 41.5 | Improved |
| add distilled residual tokenizer (ablation) | ResidualViT (reduction + interleave) · Charades-STA NLTVG ablation (ViT-L/14) | R@1 IoU=0.7: 22.8 -> 23.8 | Improved |
| add distilled residual tokenizer (ablation) | ResidualViT (reduction + interleave) · Charades-STA NLTVG ablation (ViT-L/14) | encoding cost (GFLOPs/sec): 102.0 -> 102.6 | No clear effect |
| apply token reduction (85% dropped) to all frames vs full CLIP (ablation) | CLIP ViT-L/14 · Charades-STA NLTVG ablation (ViT-L/14) | R@1 IoU=0.5: 42.9 -> 28.5 | Hurt |
| replace CLIP with ResidualViT (N=2, p=85%) | CLIP ViT-L/14 · Charades-STA zero-shot NLTVG | R@1 IoU=0.5: 42.9 -> 41.5 | No clear effect |
| replace CLIP with ResidualViT | CLIP ViT-L/14 · MAD zero-shot NLTVG | R@1 IoU=0.5: 5.0 -> 4.3 | Hurt |
| replace CLIP with ResidualViT | CLIP ViT-L/14 · ActivityNet-v1.3 TAL (ActionFormer) | mAP (%): 34.83 -> 34.46 | No clear effect |
| replace CLIP with ResidualViT | CLIP ViT-L/14 · MAD audio description | CIDEr: 11.3 -> 10.6 | Hurt |
| replace CLIP with ResidualViT | CG-DETR (B/32) · Charades-STA fully supervised CG-DETR (B/32) | R@1 IoU=0.5: 49.7 -> 48.2 | No clear effect |
| increase interleave factor N beyond 2 (ablation) | ResidualViT · Charades-STA NLTVG ablation (ViT-L/14) | R@1 accuracy | Hurt |
Closest papers by meaning
- Memory Consolidation Enables Long-Context Video Understanding (ICML 2024 · similarity 0.89)
- Re-thinking Temporal Search for Long-Form Video Understanding (CVPR 2025 · similarity 0.88)
- TimeChat: A Time-sensitive Multimodal Large Language Model for Long Video Understanding (CVPR 2024 · similarity 0.88)
- FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025 · similarity 0.88)
- EchoPrune: Interpreting Redundancy as Temporal Echoes for Efficient VideoLLMs (NeurIPS 2026 · similarity 0.88)
- Time-R1: Post-Training Large Vision Language Model for Temporal Video Grounding (NeurIPS 2025 · similarity 0.88)
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.
