Research map / Video / Video-language understanding
SViTT: Temporal Learning of Sparse Video-Text Transformers
Yi Li, Kyle Min, Subarna Tripathi, Nuno Vasconcelos
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 |
|---|---|---|---|
| Use hybrid edge and node sparsity instead of dense attention | SViTT · DiDeMo zero-shot retrieval | Recall mean: 49.3 -> 51.5 | Improved |
| Use hybrid edge and node sparsity instead of dense attention | SViTT · MSR-VTT zero-shot retrieval | Recall mean: 43.6 -> 43.8 | No clear effect |
| Use hybrid edge and node sparsity instead of dense attention | SViTT · Charades retrieval, fine-tuned | Recall mean: 32.7 -> 35.7 | Improved |
| Use hybrid edge and node sparsity instead of dense attention | SViTT · SSv2-Label retrieval, fine-tuned | Recall mean: 66.1 -> 69.3 | Improved |
| Use hybrid edge and node sparsity instead of dense attention | SViTT · ActivityNet-QA | accuracy: 42.3 -> 43.2 | Improved |
| Add text-guided token pruning at q_m=0.1 after visual pruning (ablation) | Dense video-text model · DiDeMo zero-shot retrieval | Recall mean: 46.7 -> 47.9 | Improved |
| Use edge sparsity with (1, 5, 56) instead of (1, 3, 56) (ablation) | Dense video-text model · DiDeMo retrieval | Recall mean: 37.6 -> 43.8 | Improved |
| Remove the single local diagonal-attention block (ablation) | Sparse video-text model · DiDeMo retrieval | Recall | Hurt |
| Increase the local attention window to neighboring blocks (ablation) | Sparse video-text model · DiDeMo retrieval | Recall | Hurt |
| Train using temporal sparse expansion rather than single-stage training (ablation) | SViTT · DiDeMo, 16 frames | Recall mean: 47.7 -> 52.2 | Improved |
| Reorder tokens with Morton or Hilbert curves instead of standard flattening (ablation) | SViTT · MSR-VTT retrieval | Recall mean: 38.5 -> 36.9 | Hurt |
| Evaluate with shuffled rather than normal frame order (ablation) | SViTT Hybrid · SSv2-Label retrieval | Recall: 69.3 -> 65.8 | Hurt |
Closest papers by meaning
- Revisiting Temporal Modeling for CLIP-Based Image-to-Video Knowledge Transferring (CVPR 2023 · similarity 0.90)
- Temporal Reasoning Transfer from Text to Video (ICLR 2025 · similarity 0.89)
- Less Is More: ClipBERT for Video-and-Language Learning via Sparse Sampling (CVPR 2021 · similarity 0.89)
- Not Another Text Benchmark: Putting the “Visual" Back in Visual Question Answering for Large Video Models (NeurIPS 2026 · similarity 0.88)
- TimeChat: A Time-sensitive Multimodal Large Language Model for Long Video Understanding (CVPR 2024 · similarity 0.88)
- EventLens: Event-Structure Reinforcement Learning for Video Understanding (NeurIPS 2026 · 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.
