Research map / Video / Video-language understanding
Temporal Query Networks for Fine-Grained 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 |
|---|---|---|---|
| Use query-attribute factorization with a separate learned query and classifier for each response | S3D on Diving48-v2 · Diving48-v2 | per-video accuracy: 80.4 -> 81.8 | Helped |
| Replace query-specific prediction with one shared multi-label attribute head (ablation) | S3D text-attribute classification · Diving48-v2 | per-video accuracy: 80.0 -> 50.3 | Hurt |
| Predict the attribute descriptions autoregressively in text order | S3D sequence prediction · Diving48-v2 | per-video accuracy: 80.4 -> 65.1 | Hurt |
| Add a self-attention encoder while retaining class-index classification (ablation) | S3D average-pooling baseline · Diving48-v2 | per-class accuracy: 72.3 -> 73.7 | Helped |
| Train with full temporal support instead of a fixed short frame sample (ablation) | S3D on Diving48-v2 · Diving48-v2 | per-video accuracy: 58.8 -> 80.4 | Helped |
| Continuously update feature-bank features online rather than keeping them frozen (ablation) | S3D with average pooling · Diving48-v2 | per-video accuracy: 66.6 -> 80.4 | Helped |
| Continuously update feature-bank features online rather than keeping them frozen (ablation) | S3D with TQN · Diving48-v2 | per-video accuracy: 68.2 -> 81.1 | Helped |
| Increase the warm-up maximum video length from 4 to 8 clips (ablation) | TQN two-stage training · Diving48-v2 | per-video accuracy: 78.9 -> 81.8 | Helped |
| Remove warm-up training and train directly with the feature bank (ablation) | TQN two-stage training · Diving48-v2 | per-video accuracy: 79.2 -> 81.8 | Hurt |
| Increase online feature-bank updates from 4 to 8 clips per iteration (ablation) | TQN two-stage training · Diving48-v2 | per-video accuracy: 78.3 -> 80.3 | Helped |
| Add the TQN decoder to LT-S3D with dense temporal context | S3D on Gym99 VT subset · Gym99 VT subset | per-video accuracy: 69.1 -> 74.9 | Helped |
| Pre-train TQN on Gym288 before fine-tuning on Diving48-v2 | TQN · Diving48-v2 | top-1 accuracy: 81.8 -> 83.3 | Helped |
Closest papers by meaning
- SeFAR: Semi-supervised Fine-grained Action Recognition with Temporal Perturbation and Learning Stabilization (AAAI 2025 · similarity 0.88)
- From Play to Replay: Composed Video Retrieval for Temporally Fine-Grained Videos (NeurIPS 2025 · similarity 0.87)
- SlowFocus: Enhancing Fine-grained Temporal Understanding in Video LLM (NeurIPS 2024 · similarity 0.87)
- ActionAtlas: A VideoQA Benchmark for Domain-specialized Action Recognition (NeurIPS 2024 · similarity 0.87)
- Enhancing Temporal Understanding in Video-LLMs through Stacked Temporal Attention in Vision Encoders (NeurIPS 2025 · similarity 0.86)
- Storyboard-guided Alignment for Fine-grained Video Action Recognition (NeurIPS 2025 · similarity 0.86)
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.