Research map / Detection and segmentation / Open-set and few-shot detection
Dense Siamese Network for Dense Unsupervised Learning
Wenwei Zhang, Jiangmiao Pang, Kai Chen, Chen Change Loy
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 PixSim with a 7x7 grid on top of SimSiam (ablation) | SimSiam · VOC 07+12 detection (COCO 800ep pre-train) | AP: 53.5 -> 54.9 | Improved |
| use a single sampled grid point (K=1) in PixSim (ablation) | DenseSiam · VOC 07+12 detection (COCO 800ep pre-train) | AP: 53.5 -> 28.8 | Hurt |
| raise PixSim loss weight from 0.1 to 1.0 (ablation) | DenseSiam · VOC 07+12 detection (COCO 800ep pre-train) | AP: 53.3 -> 54.9 | Improved |
| add RegionSim (weight 0.1) on top of PixSim, started at the middle of training (ablation) | DenseSiam · VOC 07+12 detection (COCO 800ep pre-train) | AP: 54.9 -> 55.5 | Improved |
| start RegionSim early at 0.4 of training instead of 0.5 (ablation) | DenseSiam · VOC 07+12 detection (COCO 800ep pre-train) | AP: 55.5 -> 37.7 | Hurt |
| add PixSim to RegionSim (ablation) | DenseSiam · VOC 07+12 detection (COCO 800ep pre-train) | AP: 52.8 -> 55.5 | Improved |
| place grid sample after projector and predictor rather than before (ablation) | DenseSiam · VOC 07+12 detection (COCO 800ep pre-train) | AP: 53.4 -> 54.9 | Improved |
| do image-level similarity on intersected regions only instead of whole image (ablation) | SimSiam · VOC 07+12 detection (COCO 800ep pre-train) | AP: 53.5 -> 0.0 | Hurt |
| build correspondence by feature similarity instead of location (ablation) | PixSim · VOC 07+12 detection (COCO 800ep pre-train) | AP: 54.9 -> 36.3 | Hurt |
| add region embeddings to location correspondence (ablation) | PixSim · VOC 07+12 detection (COCO 800ep pre-train) | AP: 54.9 -> 55.5 | Improved |
| add auxiliary over-clustering head (ablation) | PixSim segmentation · Unsupervised semantic segmentation (curated COCO) | mIoU: 10.1 -> 11.1 | Improved |
| add class-balanced CE loss (ablation) | PixSim + aux. segmentation · Unsupervised semantic segmentation (curated COCO) | mIoU: 11.1 -> 15.0 | Improved |
Closest papers by meaning
- Siamese Image Modeling for Self-Supervised Vision Representation Learning (CVPR 2023 · similarity 0.88)
- Exploring Simple Siamese Representation Learning (CVPR 2021 · similarity 0.88)
- Warp Consistency for Unsupervised Learning of Dense Correspondences (ICCV 2021 · similarity 0.87)
- DenseReg: Fully Convolutional Dense Shape Regression In-The-Wild (CVPR 2017 · similarity 0.87)
- Region-aware Distribution Contrast: A Novel Approach to Multi-Task Partially Supervised Learning (ECCV 2024 · similarity 0.87)
- Near, far: Patch-ordering enhances vision foundation models' scene understanding (ICLR 2025 · 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.
