Research map / Detection and segmentation / Weakly supervised detection and segmentation
UCOD-DPL: Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning
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 UCOD-DPL, a teacher-student unsupervised camouflaged object detector that adaptively mixes fixed-strategy and teacher pseudo-labels, uses a dual-branch adversarial decoder and a Look-Twice zoom-in refinement, reaching state-of-the-art among unsupervised methods on COD benchmarks.
Problem
Unsupervised camouflaged object detection relies on noisy, low-resolution pseudo-labels from fixed strategies and simple 1x1 conv decoders, causing poor performance especially for small objects and foreground-background confusion.
Approach
Dynamically mix fixed-strategy pseudo-labels with teacher predictions via a discriminator-based scoring function (APM), decode foreground and background with separate adversarially-regularized branches (DBA), and re-infer zoomed crops of small predicted objects (Look-Twice).
Key results
- On COD10K-Test with DINOv2, the method reaches S-measure 0.834 and weighted F-measure 0.763, versus 0.767 and 0.641 for FOUND (DINOv2) under the same unsupervised setting.
- Ablation on COD10K-Test shows the full model (Tea-Stu+APM+DBA+Look-Twice) improves S-measure from 0.626 (teacher-student only) to 0.834.
- With DINOv2 the method surpasses several semi-supervised (CamoTeacher 10%) and fully supervised methods (e.g., SINet, SINetv2) across all four test sets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the full UCOD-DPL method instead of FOUND | DINOv2-based UCOD · CHAMELEON | S-measure: .829 -> .864 | Helped |
| Use the full UCOD-DPL method instead of FOUND | DINOv2-based UCOD · CAMO-Test | S-measure: .770 -> .793 | Helped |
| Use the full UCOD-DPL method instead of FOUND | DINOv2-based UCOD · COD10K-Test | S-measure: .767 -> .834 | Helped |
| Use the full UCOD-DPL method instead of FOUND | DINOv2-based UCOD · NC4K | S-measure: .816 -> .850 | Helped |
| Add APM to teacher-student framework with a 1×1 convolution decoder (ablation) | UCOD-DPL ablation · COD10K-Test | S-measure: .626 -> .633 | Helped |
| Add APM to DBA decoder without the teacher-student framework (ablation) | UCOD-DPL ablation · COD10K-Test | S-measure: .607 -> .603 | Hurt |
| Add DBA decoder to the teacher-student framework (ablation) | UCOD-DPL ablation · COD10K-Test | S-measure: .626 -> .724 | Helped |
| Add APM to teacher-student framework with DBA decoder (ablation) | UCOD-DPL ablation · COD10K-Test | S-measure: .724 -> .788 | Helped |
| Add Look-Twice to teacher-student framework with DBA decoder (ablation) | UCOD-DPL ablation · COD10K-Test | S-measure: .724 -> .784 | Helped |
| Add Look-Twice to teacher-student framework with DBA and APM (ablation) | UCOD-DPL · COD10K-Test | S-measure: .788 -> .834 | Helped |
| Replace APM with proportional (1:1) pseudo-label mixing (ablation) | UCOD-DPL · COD10K-Test | S-measure: .834 -> .707 | Hurt |
| Replace APM with linear decay pseudo-label mixing (ablation) | UCOD-DPL · COD10K-Test | S-measure: .834 -> .809 | Hurt |
Surprising findings
- Look-Twice degrades performance when the network's base predictions are weak, as shown in ablation.
- Random Perlin noise as the fixed pseudo-label gives better results than pure black or white masks, suggesting noisy masks still provide useful signal.
Limitations
- Look-Twice may mistakenly amplify occluded or fragmented object parts as small objects.
- Performance depends on the quality of the base predictions; Look-Twice hurts weak models.
Experimental setup
Datasets: CAMO-Train COD10K-Train
Benchmarks: CHAMELEON CAMO-Test COD10K-Test NC4K
Metrics: S-measure weighted F-measure mean F-measure E-measure MAE
Baselines: FOUND UCOS-DA TokenCut SelfMask SpectralSeg BigGW CamoTeacher SCOD-ND SINet BiRefNet
Builds on: FOUND UCOS-DA BYOL DINOv2 ZoomNet
Closest papers by meaning
- Unsupervised Camouflaged Object Detection with Dual-Eigenvector Spectral Pseudo-Labeling and Contrastive Refinement (ICML 2026 · similarity 0.94)
- EReCu: Pseudo-label Evolution Fusion and Refinement with Multi-Cue Learning for Unsupervised Camouflage Detection (CVPR 2026 · similarity 0.90)
- SdalsNet: Self-Distilled Attention Localization and Shift Network for Unsupervised Camouflaged Object Detection (AAAI 2025 · similarity 0.87)
- Shift the Lens: Environment-Aware Unsupervised Camouflaged Object Detection (CVPR 2025 · similarity 0.86)
- CamoDiffusion: Camouflaged Object Detection via Conditional Diffusion Models (AAAI 2024 · similarity 0.86)
- UN-DETR: Promoting Objectness Learning via Joint Supervision for Unknown Object Detection (AAAI 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.