Research map / Detection and segmentation / Weakly supervised detection and segmentation
WISNet: Pseudo Label Generation on Unbalanced and Patch Annotated Waste Images
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces the ShanghaiWaste dataset (12,208 waste images, 40,392 patch annotations across 21 classes) and WISNet, a weakly supervised pseudo-label generation pipeline combining unsupervised subcategory classification, contrastive prototype feature learning, and iterative patch expansion, reaching 40.2% mIoU with PANet.
Problem
Semantic segmentation of densely overlapped, stained, and class-imbalanced waste images for sorting assessment, where only coarse rectangular patch-level annotations are available and pixel-level labels are costly.
Approach
Generate pixel-level pseudo labels from rectangular patch annotations by training a feature extractor with contrastive prototype learning supervised additionally by K-Means subcategory labels, then iteratively expanding patches via prototype-based cosine assignment until convergence.
Key results
- WISNet with PANet achieves 40.2% mIoU on ShanghaiWaste-Seg, versus 34.0% for DuPL-based pseudo labels with PANet under the same setup (second-best ACR/PPC around 36%).
- WISNet with DeepLab reaches 31.8% mIoU, exceeding the second-best methods (ACR and PPC) by 2.7 points.
- Ablation: PUC adds +2.0 mIoU and IPE adds +3.8 mIoU over the PANet baseline (36.4%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use WISNet-generated pseudo labels instead of PPC-generated labels | PANet · ShanghaiWaste-Seg | mIoU: 36.1 -> 40.2 | Helped |
| use WISNet-generated pseudo labels instead of ACR-generated labels | DeepLab · ShanghaiWaste-Seg | mIoU: 29.1 -> 31.8 | Helped |
| add Pre-trained Unsupervised Classification (PUC) (ablation) | PANet baseline · ShanghaiWaste-Seg | mIoU: 36.4 -> 38.4(+2.0) | Helped |
| add Iterative Patch Generation (IPG) to PUC (ablation) | PANet + PUC · ShanghaiWaste-Seg | mIoU: 38.4 -> 39.5(+3.1) | Helped |
| add Iterative Patch Expansion (IPE) (ablation) | PANet + PUC + IPG · ShanghaiWaste-Seg | mIoU: 39.5 -> 40.2(+3.8) | Helped |
| replace manual patch-level annotations with IPE-generated labels | DeepLab · ShanghaiWaste-Seg | mIoU: 23.0 -> 31.8 | Helped |
| replace manual patch-level annotations with IPE-generated labels | PANet · ShanghaiWaste-Seg | mIoU: 36.4 -> 40.2 | Helped |
| add IPE refinement to PSA-generated labels (ablation) | DeepLab · ShanghaiWaste-Seg | mIoU: 28.9 -> 30.1 | Helped |
| add IPE refinement to IRN-generated labels (ablation) | DeepLab · ShanghaiWaste-Seg | mIoU: 30.1 -> 31.3 | Helped |
| add IPE refinement to PSA-generated labels (ablation) | PANet · ShanghaiWaste-Seg | mIoU: 34.4 -> 36.1 | Helped |
| add IPE refinement to IRN-generated labels (ablation) | PANet · ShanghaiWaste-Seg | mIoU: 36.6 -> 37.6 | Helped |
| reduce patch size to αS (ablation) | WISNet · ShanghaiWaste-Seg | mIoU: 40.2 -> 36.6; 40.2 -> 34.1 | Hurt |
Surprising findings
- ASGnet, which achieves SOTA on fully supervised VOC/COCO, performs worst among the three segmentation models (20.4% mIoU) under patch supervision.
- Synthetically shrinking patches to a single point still yields usable segmentation (32.4% mIoU), suggesting robustness to annotation quality.
Limitations
- Evaluation is on a relatively small pixel-labeled test set of 928 images from two residential complexes.
- Relies on a pretrained ImageNet feature extractor and K-Means for subcategory labels; K is fixed to 3.
- Performance on rare categories (e.g., leather, paint bucket) remains very low.
Experimental setup
Datasets: ShanghaiWaste ShanghaiWaste-Seg
Benchmarks: ShanghaiWaste-Seg
Metrics: mIoU
Baselines: SDI BANA BBAM AEFT PPC ACR DuPL DeepLab V3 ASGnet PSA IRN
Builds on: PANet DuPL PSA IRN
Closest papers by meaning
- UrbanWaste: In-the-Bin Dataset for Waste Disposal Inspection with Multi-Granularity Hierarchical Labels (AAAI 2025 · similarity 0.88)
- Label-Efficient Few-Shot Semantic Segmentation with Unsupervised Meta-Training (AAAI 2024 · similarity 0.82)
- W2P: Switching from Weak Supervision to Partial Supervision for Semantic Segmentation (AAAI 2024 · similarity 0.82)
- RWMS: Reliable Weighted Multi-Phase for Semi-supervised Segmentation (AAAI 2024 · similarity 0.81)
- Towards Robust Pseudo-Label Learning in Semantic Segmentation: An Encoding Perspective (NeurIPS 2025 · similarity 0.81)
- AllSpark: Reborn Labeled Features from Unlabeled in Transformer for Semi-Supervised Semantic Segmentation (CVPR 2024 · similarity 0.81)
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.