Research map / Detection and segmentation / Weakly supervised detection and segmentation
Looking Beyond Single Images for Contrastive Semantic Segmentation Learning
Read the paperCheck 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 |
|---|---|---|---|
| add cross-image contrastive learning with auxiliary labels vs ImageNet pre-training | DeepLabV3 ResNet-50 · NYUv2 10% labels | mIoU (%): 20.4 -> 29.8 | Helped |
| use auxiliary-label contrastive pretraining instead of PseudoSeg semi-supervised | DeepLabV3 · NYUv2 100% | mIoU (%): 47.1 -> 49.2 | Helped |
| generate auxiliary labels with DeepLabV3 trained on target dataset instead of MoCo ResNet-50 (ablation) | auxiliary label generation · NYUv2 10% labels | mIoU (%): 29.8 -> 36.1 | Helped |
| use unsupervised MoCo feature extractor instead of ImageNet-supervised ResNet-50 (ablation) | auxiliary label generation · NYUv2 10% labels | mIoU (%): 30.4 -> 29.8 | No clear effect |
| add superpixel voting refinement to auxiliary labels (ablation) | auxiliary labels · NYUv2 / Cityscapes | mIoU (%): +0.7-1.1% | Helped |
| use DBSCAN instead of K-means clustering (ablation) | auxiliary labels · NYUv2 / Cityscapes | mIoU (%) | Hurt |
| increase pairs per iteration from 1k/4k to 10k/40k anchors/contrast (ablation) | contrastive training · NYUv2 | mIoU (%): 42.1 -> 42.7 | Helped |
| add momentum memory bank (ablation) | contrastive training · NYUv2 | mIoU (%): 43.5 -> 43.7 | Helped |
| add multi-scale contrastive loss (ablation) | contrastive training · NYUv2 | mIoU (%): 45.0 -> 45.2 | Helped |
| add auxiliary contrastive loss on mid-level features (ablation) | contrastive training · NYUv2 | mIoU (%): 45.2 -> 45.6 | Helped |
| add confidence weighting of contrastive loss (ablation) | contrastive training · NYUv2 | mIoU (%): 45.6 -> 46.0 | Helped |
| increase crop size from 321x361 to 473x473 (smaller batch) (ablation) | contrastive training · NYUv2 | mIoU (%): 43.5 -> 43.3 | Hurt |
Closest papers by meaning
- Boosting Semi-Supervised Semantic Segmentation with Probabilistic Representations (AAAI 2023 · similarity 0.90)
- Bootstrapping Semantic Segmentation with Regional Contrast (ICLR 2022 · similarity 0.89)
- Revisiting Contrastive Methods for Unsupervised Learning of Visual Representations (NeurIPS 2021 · similarity 0.88)
- Regional Semantic Contrast and Aggregation for Weakly Supervised Semantic Segmentation (CVPR 2022 · similarity 0.88)
- CIAN: Cross-Image Affinity Net for Weakly Supervised Semantic Segmentation (AAAI 2020 · similarity 0.88)
- Learning Hierarchical Image Segmentation For Recognition and By Recognition (ICLR 2024 · 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.