Research map / Video / Action recognition
DistInit: Learning Video Representations Without a Single Labeled Video
Rohit Girdhar, Du Tran, Lorenzo Torresani, Deva Ramanan
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 |
|---|---|---|---|
| initialize with DistInit instead of scratch | Res3D-18 · HMDB-51 split 1 | top-1 accuracy: 25.4 -> 39.9 | Improved |
| initialize with DistInit instead of ImageNet inflation | Res3D-18 · HMDB-51 split 1 | top-1 accuracy: 35.8 -> 39.9 | Improved |
| initialize with DistInit instead of scratch | R(2+1)D-18 · HMDB-51 split 1 | top-1 accuracy: 24.1 -> 40.3 | Improved |
| pick a random frame for teacher target instead of the center frame (ablation) | DistInit R(2+1)D-18 · HMDB-51 split 1 | top-1 accuracy: 40.3 -> 43.2 | Improved |
| average teacher logits over 2 random frames instead of the center frame (ablation) | DistInit R(2+1)D-18 · HMDB-51 split 1 | top-1 accuracy: 40.3 -> 44.0 | Improved |
| use cross-entropy over softmax instead of MSE over logits (ablation) | DistInit R(2+1)D-18 · HMDB-51 split 1 | top-1 accuracy: 39.9 -> 40.3 | No clear effect |
| use Kinetics instead of Sports1M unlabeled videos (ablation) | DistInit R(2+1)D-18 · HMDB-51 split 1 | top-1 accuracy: 39.9 -> 40.3 | No clear effect |
| combine Kinetics and Sports1M unlabeled videos (ablation) | DistInit R(2+1)D-18 · HMDB-51 split 1 | top-1 accuracy: 40.3 -> 41.8 | Improved |
| use ImageNet teacher instead of PlaceNet teacher (ablation) | DistInit R(2+1)D-18 · HMDB-51 split 1 | top-1 accuracy: 36.8 -> 40.3 | Improved |
| add PlaceNet teacher to the ImageNet teacher (ablation) | DistInit R(2+1)D-18 · HMDB-51 split 1 | top-1 accuracy: 40.3 -> 40.7 | No clear effect |
| drop teacher low-confidence videos (entropy above 2.7, about 50% of data) (ablation) | DistInit R(2+1)D-18 · HMDB-51 split 1 | top-1 accuracy | Improved |
| initialize flow network with DistInit instead of ImageNet mean inflation | Res3D-18 / R(2+1)D-18 on optical flow · HMDB-51 split 1 | top-1 accuracy: 43.9 -> 49.2 | Improved |
Later papers that compared with this method
1 later accepted paper reports DistInit as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Motion-Augmented Self-Training for Video Recognition at Smaller Scale (ICCV 2021 · 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
- Masked Video Distillation: Rethinking Masked Feature Modeling for Self-Supervised Video Representation Learning (CVPR 2023 · similarity 0.89)
- Disentangling Spatial and Temporal Learning for Efficient Image-to-Video Transfer Learning (ICCV 2023 · similarity 0.88)
- Improving Video Model Transfer With Dynamic Representation Learning (CVPR 2022 · similarity 0.88)
- Dancing with Still Images: Video Distillation via Static-Dynamic Disentanglement (CVPR 2024 · similarity 0.88)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.87)
- 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.
