Research map / Video / Action recognition
UnweaveNet: Unweaving Activity Stories
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 |
|---|---|---|---|
| pretrain on synthetic stories before finetuning on activity stories rather than train on activity stories only | UnweaveNet · EPIC-KITCHENS activity-story test | RI: 67.0 -> 75.1 | Helped |
| pretrain on synthetic stories before finetuning on activity stories rather than train on activity stories only | UnweaveNet · EPIC-KITCHENS activity-story test | TFA: 57.3 -> 77.5 | Helped |
| pretrain only on synthetic stories instead of activity stories only | UnweaveNet · EPIC-KITCHENS activity-story test | RI: 67.0 -> 69.1 | Helped |
| pretrain only on synthetic stories instead of activity stories only | UnweaveNet · EPIC-KITCHENS activity-story test | ΔN: 0.3 -> 2.0 | Hurt |
| use a randomized 2--4s gap instead of a fixed 2s gap within synthetic threads (ablation) | UnweaveNet synthetic-story pretraining · EPIC-KITCHENS synthetic-story pretraining ablation | RI: 74.6$\pm$0.7 -> 75.1$\pm 0.5 | Helped |
| increase the maximum synthetic threads from 2 to 4 (ablation) | UnweaveNet synthetic-story pretraining · EPIC-KITCHENS synthetic-story pretraining ablation | RI: 69.0$\pm$0.6 -> \textbf{75.1}$\pm$0.5 | Helped |
| increase the maximum synthetic threads from 4 to 6 (ablation) | UnweaveNet synthetic-story pretraining · EPIC-KITCHENS synthetic-story pretraining ablation | RI: \textbf{75.1}$\pm$0.5 -> 73.5$\pm$0.8 | Hurt |
| use a transformer controller instead of linear embedding, with GRU thread updates (ablation) | UnweaveNet · EPIC-KITCHENS architecture ablation | RI: 74.3$\pm$0.3 -> \textbf{75.1}$\pm$0.5 | Helped |
| use GRU thread updates instead of replacing the thread representation with the last clip, with transformer controller (ablation) | UnweaveNet · EPIC-KITCHENS architecture ablation | RI: 74.8$\pm$0.3 -> \textbf{75.1}$\pm$0.5 | Helped |
| use non-uniform scenario loss weights (1, 100, 10) instead of equal weights (ablation) | UnweaveNet · EPIC-KITCHENS activity-story test | RI: 67.9\% -> 75.1\% | Helped |
| increase weights for resume and new-thread decisions (ablation) | UnweaveNet · EPIC-KITCHENS activity-story test | continue-thread performance | Hurt |
| add random crops or color distortion as data augmentation (ablation) | UnweaveNet · EPIC-KITCHENS activity-story test | performance | No clear effect |
Closest papers by meaning
- Learning Goals From Failure (CVPR 2021 · similarity 0.87)
- Oops! Predicting Unintentional Action in Video (CVPR 2020 · similarity 0.85)
- Video PreTraining (VPT): Learning to Act by Watching Unlabeled Online Videos (NeurIPS 2022 · similarity 0.84)
- Watch and Learn: Learning to Use Computers from Online Videos (CVPR 2026 · similarity 0.84)
- UniMTS: Unified Pre-training for Motion Time Series (NeurIPS 2024 · similarity 0.84)
- Learning to Predict Activity Progress by Self-Supervised Video Alignment (CVPR 2024 · 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.