Research map / Detection and segmentation / Weakly supervised detection and segmentation
DuPL: Dual Student with Trustworthy Progressive Learning for Robust Weakly Supervised Semantic Segmentation
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.
Proposes DuPL, a one-stage weakly supervised semantic segmentation framework with two mutually supervising student sub-networks, dynamic threshold adjustment, GMM-based adaptive noise filtering, and consistency regularization on filtered pixels.
Problem
One-stage WSSS with image-level labels suffers from CAM confirmation bias, where incorrect CAM pseudo-labels are reinforced during training, and fixed high thresholds discard many useful pixels.
Approach
A dual student co-training setup with a representation-level discrepancy loss yields diverse CAMs that supervise each other, combined with progressively lowered background thresholds, loss-distribution (GMM) noise filtering, and consistency regularization on discarded pixels so that every pixel contributes supervision.
Key results
- DuPL reaches 73.5% mIoU on VOC val and 72.8% on VOC test (ImageNet-21k ViT-B), outperforming prior one-stage methods such as ToCo (70.5 val).
- On MS COCO val, DuPL achieves 44.6% mIoU with ImageNet-21k ViT-B, versus 41.3% for ToCo with ImageNet-1k.
- CAM pseudo-label quality on VOC train reaches 75.1% mIoU with ImageNet-1k ViT-B (vs 72.2% for ToCo).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add dual student architecture with discrepancy loss (ablation) | ViT-B one-stage baseline · PASCAL VOC 2012 val | CAM mIoU: 63.2 -> 67.3 with DS + Ldis (approx. +4.1 cumulative with discrepancy) | Helped |
| add dynamic threshold adjustment (DTA) (ablation) | DS + Ldis · PASCAL VOC 2012 val | segmentation mIoU: +2.6 (64.1 -> 66.7) | Helped |
| add adaptive noise filtering (ANF) (ablation) | DS + Ldis + DTA · PASCAL VOC 2012 val | segmentation mIoU: +1.5 (66.7 -> 68.2) | Helped |
| add consistency regularization on filtered regions (Lreg) (ablation) | DS + Ldis + DTA + ANF · PASCAL VOC 2012 val | segmentation mIoU: 68.2 -> 69.9 | Helped |
| use fixed low background threshold vs dynamic threshold: dynamic threshold final tau_h(T)=0.55 vs 0.50 (ablation) | DuPL DTA · PASCAL VOC 2012 val | segmentation mIoU: 73.5 CAM / 72.2 seg at 0.55 vs 72.3 / 71.5 at 0.50 | Mixed |
| use representation-level discrepancy loss plus differently augmented inputs for dual student (ablation) | Dual student · PASCAL VOC 2012 val | segmentation mIoU: Diff. Aug + Ldis 72.2 vs Ldis-only 69.4 seg vs None 68.9; Diff. Aug alone 69.8 | Mixed |
| DuPL full method vs ToCo | ViT-B one-stage WSSS · PASCAL VOC 2012 val | segmentation mIoU: 72.2 vs 69.8 (+2.4) | Helped |
| reduce over-activation via dual student (ablation) | baseline with low threshold · PASCAL VOC 2012 val | over-activation rate: reduces OA rate by over 15% in many classes | Helped |
Surprising findings
- Dual student with only the random segmentation-head initialization gives a CAM gain of about 2% mIoU, and the two sub-nets' CAMs stay highly identical without the discrepancy loss.
Limitations
- Performance gap to fully supervised upper bound remains (90.1% of fully supervised on VOC with ImageNet-1k)
- Hyperparameters such as background thresholds and warm-up iterations are tuned on the VOC val set and sensitive to choice
Experimental setup
Datasets: PASCAL VOC 2012 SBD MS COCO 2014
Benchmarks: PASCAL VOC 2012 MS COCO 2014
Metrics: mIoU
Baselines: ToCo AFA 1Stage SLRNet TSCD ViT-PCM EPS L2G PPC ACR
Builds on: ToCo co-training patch token contrast loss Expectation Maximization
Closest papers by meaning
- Progressive Feature Self-Reinforcement for Weakly Supervised Semantic Segmentation (AAAI 2024 · similarity 0.90)
- Weakly Supervised Semantic Segmentation via Progressive Confidence Region Expansion (CVPR 2025 · similarity 0.88)
- W2P: Switching from Weak Supervision to Partial Supervision for Semantic Segmentation (AAAI 2024 · similarity 0.88)
- From SAM to CAMs: Exploring Segment Anything Model for Weakly Supervised Semantic Segmentation (CVPR 2024 · similarity 0.88)
- Separate and Conquer: Decoupling Co-occurrence via Decomposition and Representation for Weakly Supervised Semantic Segmentation (CVPR 2024 · similarity 0.88)
- SFC: Shared Feature Calibration in Weakly Supervised Semantic Segmentation (AAAI 2024 · similarity 0.87)
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.