Research map / Label-efficient and robust learning / Self-supervised and contrastive
OBoW: Online Bag-of-Visual-Words Generation for Self-Supervised Learning
Spyros Gidaris, Andrei Bursuc, Gilles Puy, Nikos Komodakis, Matthieu Cord, Patrick Perez
Check 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 a queue vocabulary of locally averaged features rather than online k-means with Sinkhorn balancing (ablation) | OBoW · ImageNet 20% linear evaluation | accuracy: 39.74 -> 45.86 | Improved |
| Use locally averaged instead of globally averaged queue features with multi-scale BoW (ablation) | OBoW · ImageNet 20% few-shot, 1-shot | accuracy: 41.29 -> 44.18 | Improved |
| Anneal teacher momentum from 0.99 to 1.0 rather than use constant low momentum (ablation) | OBoW · ImageNet 20% few-shot, 1-shot | accuracy: 12.70 -> 42.11 | Improved |
| Use dynamic weight generation instead of fixed prediction weights (ablation) | OBoW · ImageNet 20% few-shot, 1-shot | accuracy: 2.00 -> 42.11 | Improved |
| Use soft rather than hard visual-word assignments (ablation) | OBoW · ImageNet 20% few-shot, 1-shot | accuracy: 38.61 -> 42.11 | Improved |
| Use a smaller single crop (160²) rather than one 224² crop (ablation) | OBoW · ImageNet 20% linear evaluation | accuracy: 31.39 -> 45.86 | Improved |
| Add cutmix to one 224² crop (ablation) | OBoW · ImageNet 20% linear evaluation | accuracy: 31.39 -> 39.46 | Improved |
| Use two 160² crops instead of one 160² crop (ablation) | OBoW · ImageNet 20% linear evaluation | accuracy: 45.86 -> 47.64 | Improved |
| Add five 96² patches to two 160² crops (ablation) | OBoW · ImageNet 20% linear evaluation | accuracy: 47.64 -> 49.64 | Improved |
| Add multi-scale BoW targets to two 160² crops plus five 96² patches (ablation) | OBoW · ImageNet 20% linear evaluation | accuracy: 49.64 -> 50.89 | Improved |
| Use OBoW instead of BoWNet | self-supervised ResNet18 · ImageNet 20% few-shot, 1-shot | accuracy: 33.80 -> 42.11 | Improved |
| Use OBoW full rather than OBoW vanilla (ablation) | OBoW · ImageNet 20% linear evaluation | accuracy: 45.86 -> 50.89 | Improved |
Later papers that compared with this method
3 later accepted papers report OBoW as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning (ICCV 2023 · better on 1 of 1 shared result)
- VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning (ICLR 2022 · better on 0 of 3 shared results)
- Constrained Mean Shift Using Distant Yet Related Neighbors for Representation Learning (ECCV 2022 · 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
- Learning Representations by Predicting Bags of Visual Words (CVPR 2020 · similarity 0.89)
- Approximate Fisher Kernels of Non-iid Image Models for Image Categorization (TPAMI 2016 · similarity 0.87)
- Generative Object Detection with Co-Training (NeurIPS 2026 · similarity 0.87)
- OVIS: Open-Vocabulary Visual Instance Search via Visual-Semantic Aligned Representation Learning (AAAI 2022 · similarity 0.86)
- Concept-Aware Batch Sampling Improves Language-Image Pretraining (CVPR 2026 · similarity 0.86)
- T-MARS: Improving Visual Representations by Circumventing Text Feature Learning (ICLR 2024 · similarity 0.85)
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.
