Research map / Video / Action recognition
Shuffle and Learn: Unsupervised Learning Using Temporal Order Verification
Ishan Misra, C. Lawrence Zitnick, Martial Hebert
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 |
|---|---|---|---|
| add tuple-verification unsupervised pre-training instead of random init | Tuple verification · UCF101 | mean accuracy: 38.6 -> 50.2 | Improved |
| add tuple-verification unsupervised pre-training instead of random init | Tuple verification · HMDB51 | mean accuracy: 13.3 -> 18.1 | Improved |
| use a larger positive window tau_max 60 instead of 30 (ablation) | Tuple verification · UCF101 split 1 | action recognition accuracy: 47.2 -> 50.9 | Improved |
| use a small negative distance tau_min 15 instead of 60 (ablation) | Tuple verification · UCF101 split 1 | action recognition accuracy: 49.1 -> 50.9 | Improved |
| use class ratio 0.75 neg / 0.25 pos instead of balanced 0.5/0.5 (ablation) | Tuple verification · UCF101 split 1 | action recognition accuracy: 38.1 -> 50.9 | Improved |
| raise negative ratio from 0.75 to 0.85 (ablation) | Tuple verification · UCF101 split 1 | action recognition accuracy: 50.9 -> 48.6 | Hurt |
| use three-frame ordering instead of two-frame ordering (ablation) | Tuple verification · UCF101 split 1 | accuracy: 44.1 -> 50.9 | Improved |
| use three-frame order verification instead of DrLim temporal smoothness | Tuple verification · UCF101 split 1 | accuracy: 45.7 -> 50.9 | Improved |
| add tuple pre-training on top of ImageNet initialization | Tuple verification · HMDB51 | mean accuracy: 28.5 -> 29.9 | Improved |
| use tuple verification init instead of random init | Tuple verification · FLIC | mean PCK: 74.5 -> 84.7 | Improved |
| use tuple verification init instead of UCF supervised init | Tuple verification · FLIC | mean PCK: +7.6% | Improved |
| use tuple verification init instead of UCF supervised init | Tuple verification · MPII | PCKh@0.5 Full: +2.1% | Improved |
Closest papers by meaning
- Unsupervised Representation Learning by Sorting Sequences (ICCV 2017 · similarity 0.91)
- Learning Actionness via Long-range Temporal Order Verification (ECCV 2020 · similarity 0.88)
- Learning Video Representations by Transforming Time (ECCV 2020 · similarity 0.88)
- Long-Term Temporal Convolutions for Action Recognition (TPAMI 2017 · similarity 0.88)
- Slow and Steady Feature Analysis: Higher Order Temporal Coherence in Video (CVPR 2016 · similarity 0.87)
- Self-Supervised Learning of Pose Embeddings From Spatiotemporal Relations in Videos (ICCV 2017 · 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.
