Research map / Video / Action recognition
Motion-Augmented Self-Training for Video Recognition at Smaller Scale
Kirill Gavrilyuk, Mihir Jain, Ilia Karmanov, Cees G. M. Snoek
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 |
|---|---|---|---|
| use motion (optical flow) representation instead of appearance for the pseudo-label generator (ablation) | M_MPLG · UCF101 split 1 | top-1 accuracy: 59.1 -> 82.6 | Improved |
| add multi-clip training (R=3) instead of single clip (ablation) | M_MPLG motion · UCF101 split 1 | top-1 accuracy: 78.2 -> 82.6 | Improved |
| use 3 clips rather than longer 32-frame input (ablation) | M_MPLG motion · UCF101 split 1 | top-1 accuracy: 80.8 -> 82.6 | Improved |
| add multi-clip training to appearance representation (ablation) | M_MPLG appearance · UCF101 split 1 | top-1 accuracy: 58.9 -> 59.1 | No clear effect |
| add clip-order prediction term (lambda<1) to motion generator (ablation) | M_MPLG motion · UCF101 split 1 | top-1 accuracy | No clear effect |
| use clip-level instead of video-level pseudo-labels (ablation) | MotionFit self-training · UCF101 split 1 | top-1 accuracy: 76.5 -> 80.3 | Improved |
| use clip-level instead of segment-level pseudo-labels (ablation) | MotionFit self-training · UCF101 split 1 | top-1 accuracy: 79.0 -> 80.3 | Improved |
| add next/previous clip pseudo-label prediction (ablation) | MotionFit self-training · UCF101 split 1 | top-1 accuracy: +0.4% | Improved |
| increase number of clusters from 1000 to 1600 (ablation) | MotionFit (Kinetics-val) · UCF101 split 1 | top-1 accuracy: 79.7 -> 74.2 | Hurt |
| increase number of clusters from 128 to 1000 (ablation) | MotionFit (Kinetics-train) · UCF101 split 1 | top-1 accuracy: 85.2 -> 86.5 | Improved |
| fine-tune with 32 frames instead of 16 | MotionFit · UCF101 | top-1 accuracy: 87.4 -> 88.9 | Improved |
| fine-tune with 32 frames instead of 16 | MotionFit · HMDB51 | top-1 accuracy: 56.4 -> 61.4 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- MARS — MARS: Motion-Augmented RGB Stream for Action Recognition (CVPR 2019) · better on 1 of 1 shared result
- DistInit — DistInit: Learning Video Representations Without a Single Labeled Video (ICCV 2019) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Self-Supervised Video Representation Learning by Context and Motion Decoupling (CVPR 2021 · similarity 0.89)
- Self-Supervised Learning of Compressed Video Representations (ICLR 2021 · similarity 0.88)
- Self-Supervised Motion Learning From Static Images (CVPR 2021 · similarity 0.88)
- TTVidT: Decoupling the Temporal Axis for Efficient Motion-Centric Video Pretraining (NeurIPS 2026 · similarity 0.86)
- No More Shortcuts: Realizing the Potential of Temporal Self-Supervision (AAAI 2024 · similarity 0.86)
- Is ImageNet worth 1 video? Learning strong image encoders from 1 long unlabelled video (ICLR 2024 · 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.
