Research map / Video / Video-language understanding
Long-Term Feature Banks for Detailed Video Understanding
Check 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 long-term feature-bank context (ablation) | 3D CNN · AVA val | mAP: 22.1 -> 25.5 | Helped |
| Extend temporal support by increasing temporal stride (ablation) | 3D CNN · AVA val | mAP: 22.1 -> 15.7 | Hurt |
| Use past-only causal context instead of batch context (ablation) | LFB · AVA val | mAP: 25.5 -> 24.2 | Hurt |
| Use decoupled, task-adapted short-term features rather than one fixed feature representation (ablation) | LFB · AVA val | mAP: 24.3 -> 25.5 | Helped |
| Replace non-local attention with average pooling (ablation) | LFB · AVA val | mAP: 25.5 -> 23.1 | Hurt |
| Use detected actor features rather than global or grid features (ablation) | LFB · AVA val | mAP: 25.1 -> 25.5 | Helped |
| Add a second non-local layer (ablation) | LFB · AVA val | mAP: 25.1 -> 25.5 | Helped |
| Remove layer normalization (ablation) | LFB 2L · AVA val | mAP: 25.5 -> 23.9 | Hurt |
| Add a third non-local layer (ablation) | LFB 2L · AVA val | mAP: 25.5 -> 25.8 | Helped |
| Use Max rather than non-local attention for the feature-bank operator (ablation) | LFB on EPIC-Kitchens · EPIC-Kitchens validation, nouns | top-1 accuracy: 29.2 -> 31.8 | Helped |
| Add object-centric long-term features (ablation) | 3D CNN noun classifier · EPIC-Kitchens validation, nouns | top-1 accuracy: 26.1 -> 31.8 | Helped |
| Use the shorter training schedule (ablation) | 3D CNN · Charades trainval | mAP: 33.8 -> 38.3 | Helped |
Closest papers by meaning
- ReWind: Understanding Long Videos with Instructed Learnable Memory (CVPR 2025 · similarity 0.84)
- MA-LMM: Memory-Augmented Large Multimodal Model for Long-Term Video Understanding (CVPR 2024 · similarity 0.83)
- Enhancing Temporal Understanding in Video-LLMs through Stacked Temporal Attention in Vision Encoders (NeurIPS 2025 · similarity 0.83)
- Re-thinking Temporal Search for Long-Form Video Understanding (CVPR 2025 · similarity 0.83)
- Leveraging Textual Memory and Key Frame Reasoning for Full Video Understanding Using Off-the-Shelf LLMs and VLMs (Student Abstract) (AAAI 2025 · similarity 0.82)
- VideoLucy: Deep Memory Backtracking for Long Video Understanding (NeurIPS 2025 · similarity 0.82)
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.