Research map / Label-efficient and robust learning / Self-supervised and contrastive
Colorization as a Proxy Task for Visual Understanding
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| replace L*a*b regression with hue/chroma histogram prediction (ablation) | Colorization pretraining with VGG-16 · VOC 2012 Segmentation val | mIU: 48.0 -> 52.9 | Helped |
| omit hypercolumns during colorization pretraining (ablation) | Histogram colorization with VGG-16 · VOC 2012 Segmentation val | mIU: 52.9 -> 52.7 | No clear effect |
| add a larger field of view (ablation) | Colorization-pretrained VGG-16 · VOC 2012 Segmentation val | mIU: 50.7 -> 52.9 | Helped |
| use ResNet-152 instead of AlexNet or VGG-16 (ablation) | Colorization pretraining · ImNt-100k | top-5 accuracy: 48.2 -> 59.0 -> 63.1 | Helped |
| extend colorization pretraining from 3 to 10 epochs (ablation) | Colorization-pretrained VGG-16 · VOC 2012 Segmentation | mIU: 52.9% to 56.0% | Helped |
| omit learning-rate drops during colorization pretraining (ablation) | Colorization-pretrained VGG-16 · VOC 2012 Segmentation | mIU: 49.1% compared to 52.9% | Hurt |
| re-introduce color input after grayscale colorization pretraining (ablation) | Colorization-pretrained VGG-16 · VOC 2012 Segmentation | mIU: 56.0 -> 55.9 | No clear effect |
| reduce ImageNet pretraining data from 1.3M to 100k samples (ablation) | ImageNet classification pretraining · VOC 2012 Segmentation val-gray | mIU: 62.0 -> 57.1 | Hurt |
| reduce ImageNet label space from 1000 classes to 16 hierarchical buckets (ablation) | ImageNet classification pretraining · VOC 2012 Segmentation val-gray | mIU: 62.0 -> 60.0 | Hurt |
| reduce ImageNet label space from 1000 classes to 2 hierarchical buckets (ablation) | ImageNet classification pretraining · VOC 2012 Segmentation val-gray | mIU: 62.0 -> 46.1 | Hurt |
| add 10% random label noise (ablation) | ImageNet classification pretraining · VOC 2012 Segmentation val-gray | mIU: 62.0 -> 61.8 | No clear effect |
| add 50% random label noise (ablation) | ImageNet classification pretraining · VOC 2012 Segmentation val-gray | mIU: incurs a 2.6-point penalty | Hurt |
Closest papers by meaning
- Mix-and-Match Tuning for Self-Supervised Semantic Segmentation (AAAI 2018 · similarity 0.88)
- How Well Do Self-Supervised Models Transfer? (CVPR 2021 · similarity 0.88)
- How Useful Is Self-Supervised Pretraining for Visual Tasks? (CVPR 2020 · similarity 0.88)
- When Does Contrastive Visual Representation Learning Work? (CVPR 2022 · similarity 0.88)
- Visualization of Supervised and Self-Supervised Neural Networks via Attribution Guided Factorization (AAAI 2021 · 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.