Research map / Detection and segmentation / CNN features
Propagate Yourself: Exploring Pixel-Level Consistency for Unsupervised Visual Representation Learning
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 instance-level pre-training with pixel-to-propagation consistency | ResNet-50 representations · Pascal VOC, COCO, and Cityscapes dense prediction | AP, mAP, and mIoU: 2.6 AP, 0.8 / 1.0 mAP, and 1.0 mIoU | Helped |
| add pixel propagation module to PixPro (ablation) | PixPro · Pascal VOC and COCO | AP and mAP: 0.8 AP; 1.1 mAP | Helped |
| add pixel propagation module to PixContrast (ablation) | PixContrast · Pascal VOC and COCO | AP and mAP: 58.1 -> 52.9 AP; 38.8 -> 37.5 mAP | Hurt |
| use pixel-to-propagation consistency instead of pixel contrast (ablation) | pixel-level pre-training · Pascal VOC and COCO | AP and mAP: 0.7 AP; 2.0 mAP | Helped |
| use a distance threshold other than T=0.7 (ablation) | PixPro with C5 features · Pascal VOC | AP: 58.8 -> 58.3; 58.8 -> 56.8; 58.8 -> 56.5 | Hurt |
| vary the similarity sharpness exponent away from gamma=2 (ablation) | PixPro · Pascal VOC and COCO | AP and mAP: gamma=0.5: 57.9 AP, 39.7 mAP; gamma=4: 58.0 AP, 40.0 mAP; default: 58.8 AP, 40.8 mAP | Hurt |
| use zero transformation layers in g instead of one (ablation) | PixPro · Pascal VOC and COCO | AP and mAP: 58.8 -> 58.6 AP; 40.8 -> 39.4 mAP | Hurt |
| use two transformation layers in g instead of one (ablation) | PixPro · Pascal VOC and COCO | AP and mAP: 58.8 -> 58.9 AP; 40.8 -> 40.3 mAP | Mixed |
| use P3-P6 feature maps instead of C5 (ablation) | PixPro pre-training · Pascal VOC and COCO | AP and mAP: 58.8 -> 55.8 AP; 40.8 -> 41.3 mAP | Mixed |
| increase pre-training from 100 to 400 epochs (ablation) | PixPro · Pascal VOC and COCO | AP and mAP: 58.8 -> 60.2 AP; 40.8 -> 41.0 mAP | Helped |
| add instance-level SimCLR* pre-training to PixPro (ablation) | PixPro · ImageNet-1K linear evaluation and dense transfer | top-1 accuracy and AP/mAP: 55.1 -> 66.3 top-1; 58.8 -> 58.7 VOC AP; 40.8 -> 40.9 COCO mAP | Mixed |
| pre-train FPN and head layers in addition to the backbone (ablation) | PixPro · COCO (FCOS) | mAP: 37.8 -> 38.6 | Helped |
Closest papers by meaning
- Revisiting Contrastive Methods for Unsupervised Learning of Visual Representations (NeurIPS 2021 · similarity 0.91)
- Beyond Supervised vs. Unsupervised: Representative Benchmarking and Analysis of Image Representation Learning (CVPR 2022 · similarity 0.89)
- When Does Contrastive Visual Representation Learning Work? (CVPR 2022 · similarity 0.89)
- Dense Contrastive Learning for Self-Supervised Visual Pre-Training (CVPR 2021 · similarity 0.89)
- A Simple Framework for Contrastive Learning of Visual Representations (ICML 2020 · similarity 0.89)
- Perceptual Inductive Bias Is What You Need Before Contrastive Learning (CVPR 2025 · similarity 0.88)
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.