Research map / Detection and segmentation / Open-set and few-shot detection
Distilling Localization for Self-Supervised Representation Learning
Nanxuan Zhao, Zhirong Wu, Rynson W.H. Lau, Stephen Lin
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 saliency-guided copy-and-paste background augmentation | InstDist, CMC, MoCo-v1, and MoCo-v2 · ImageNet validation | linear accuracy: 56.5 -> 59.3 and 62.9; 63.4 -> 65.0 and 66.9; 60.6 -> 62.8 and 65.0; 67.5 -> 67.9 and 69.2 | Improved |
| Use class-agnostic saliency masks for copy-and-paste augmentation (ablation) | MoCo · ImageNet validation | linear accuracy: 60.6 -> 62.7; 60.6 -> 62.1; 60.6 -> 62.8; 60.6 -> 65.0 | Improved |
| Apply copy-and-paste augmentation to 30%, 50%, or 70% of images rather than none (ablation) | MoCo · ImageNet validation | linear accuracy: 60.6 -> 62.8; 60.6 -> 62.2; 60.6 -> 61.6 | Improved |
| Apply copy-and-paste augmentation to every image (ablation) | MoCo · ImageNet validation | linear accuracy: 60.6 -> 47.6 | Hurt |
| Replace backgrounds with texture images (ablation) | MoCo with DiLo · ImageNet validation | linear accuracy: 60.6 -> 60.6 (+0.0) | No clear effect |
| Replace backgrounds with ImageNet crops (ablation) | MoCo with DiLo · ImageNet validation | linear accuracy: 60.6 -> 62.1 | Improved |
| Replace backgrounds with homogeneous grayscale images (ablation) | MoCo with DiLo · ImageNet validation | linear accuracy: 60.6 -> 62.8 | Improved |
| Use Gaussian blending or mixed blending when pasting (ablation) | MoCo with DiLo · ImageNet validation | linear accuracy: 60.6 -> 62.5; 60.6 -> 62.8 | Improved |
| Mask pooled final-layer features using saliency instead of training with copy-and-paste augmentations (ablation) | MoCo · ImageNet validation | linear accuracy: -19% | Hurt |
| Add intense color jittering (ablation) | MoCo self-supervised representation · ImageNet validation | linear accuracy: 40.4 -> 56.9 | Improved |
| Add intense color jittering (ablation) | Supervised representation · ImageNet validation | linear accuracy: 77.5 -> 77.4 | Hurt |
| Add RBD or BASNet background augmentation | MoCo · VOC 2007 test | AP: 55.9 -> 56.5; 55.9 -> 56.9 | Improved |
Closest papers by meaning
- Revisiting Contrastive Methods for Unsupervised Learning of Visual Representations (NeurIPS 2021 · similarity 0.89)
- Demystifying Contrastive Self-Supervised Learning: Invariances, Augmentations and Dataset Biases (NeurIPS 2020 · similarity 0.89)
- Unsupervised Object Localization: Observing the Background To Discover Objects (CVPR 2023 · similarity 0.88)
- When Does Contrastive Visual Representation Learning Work? (CVPR 2022 · similarity 0.88)
- Propagate Yourself: Exploring Pixel-Level Consistency for Unsupervised Visual Representation Learning (CVPR 2021 · similarity 0.87)
- Unsupervised Salient Instance Detection (CVPR 2024 · 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.
