Research map / Video / Action recognition
Beyond Short Clips: End-to-End Video-Level Learning With Collaborative Memories
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 |
|---|---|---|---|
| Train with multiple sampled clips and collaborative memory rather than one clip | Slow-only · Kinetics-400 | Top-1 accuracy: 74.5% vs 77.1% | Helped |
| Train with multiple sampled clips and collaborative memory rather than one clip | Slow-only with 8×2 clips · Kinetics-400 | Top-1 accuracy: 73.2% vs 76.4% | Helped |
| Remove information sharing across the sampled clips (ablation) | Video-level training on Slow-only · Kinetics-400 | Top-1 accuracy: 75.5 vs 76.8 | Hurt |
| Allow end-to-end updating of the backbone (ablation) | Collaborative Memory on Slow-only · Kinetics-400 | Top-1 accuracy: 75.9% vs 76.8% | Helped |
| Replace associative memory with average pooling (ablation) | Collaborative Memory · Kinetics-400 | Top-1 accuracy: 75.8 vs 76.8 | Hurt |
| Replace feature gating with residual context infusion (ablation) | Collaborative Memory · Kinetics-400 | Top-1 accuracy: 76.0 vs 76.8 | Hurt |
| Increase channel reduction ratio to α=8 (ablation) | Collaborative Memory · Kinetics-400 | Top-1 accuracy: 76.8 -> 76.4 | Hurt |
| Train the backbone and memory from scratch instead of stage-wise (ablation) | Slow-only · Kinetics-400 | Top-1 accuracy: 76.8 -> 76.1 | Hurt |
| Train the backbone and memory from scratch instead of stage-wise (ablation) | R(2+1)D · Kinetics-400 | Top-1 accuracy: 78.0 -> 77.7 | Hurt |
| Use multi-iteration instead of batch reduction (ablation) | Collaborative Memory, Slow-only and R(2+1)D · Kinetics-400 | Top-1 accuracy: 76.8 vs 76.6; 78.0 vs 77.9 | No clear effect |
| Increase temporal stride alone to extend coverage (ablation) | Clip-based video backbones · Kinetics-400 | Top-1 accuracy: Slow-only: 73.2, 74.3, 74.4, 74.4; R(2+1)D: 75.7, 76.4, 75.0, 72.2 | Mixed |
| Increase the number of clips from 1 to 9 (ablation) | Collaborative Memory · AVA v2.2 | mAP: 26.9 -> 29.8 | Helped |
Closest papers by meaning
- An Empirical Study of End-to-End Temporal Action Detection (CVPR 2022 · similarity 0.89)
- Learning End-to-end Video Classification with Rank-Pooling (ICML 2016 · similarity 0.89)
- Rep Deep & Machine Learning: Exemplar-Free Continual Video Action Recognition via Slow-Fast Collaborative Learning (AAAI 2026 · similarity 0.88)
- End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames (CVPR 2024 · similarity 0.87)
- No More Shortcuts: Realizing the Potential of Temporal Self-Supervision (AAAI 2024 · similarity 0.87)
- TTVidT: Decoupling the Temporal Axis for Efficient Motion-Centric Video Pretraining (NeurIPS 2026 · 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.