Research map / Detection and segmentation / Open-set and few-shot detection
SdalsNet: Self-Distilled Attention Localization and Shift Network for Unsupervised Camouflaged Object Detection
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 SdalsNet, a fully unsupervised, non-end-to-end camouflaged object detector that localizes attention centers in a DINO-style ViT self-distillation framework and binarizes the attention map for segmentation.
Problem
Unsupervised camouflaged object detection (UCOD) without human labels, where end-to-end models pretrained on ImageNet classification struggle to precisely localize camouflaged objects.
Approach
Instead of end-to-end pixel segmentation, locate the attention center via a distance-weighted attention map (foreground argmin, background argmax), crop local regions around these centers, and train a ViT teacher-student with pull/push losses to shift attention toward the object; the final mask is obtained by thresholding and average-pool denoising of the attention map.
Key results
- On CHAMELEON, SdalsNet reports S-alpha 0.724, E-max 0.837, F-max 0.666, MAE 0.08, outperforming all listed UCOD baselines including UCOD-DA (0.715 S-alpha).
- On COD10K, SdalsNet reports S-alpha 0.696 and F-max 0.572, versus UCOD-DA's 0.689 and 0.548.
- Ablation shows removing the Frozen Training stage or Class-Head degrades performance; removing Lf causes the largest drop (about 29.87% in F-max on the reported comparison).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the Attention Localization Module (ALM) (ablation) | SdalsNet · CAMO | Fmaxβ: 0.525 -> 0.661 | Helped |
| add Attention Shift Learning (ASL) (ablation) | SdalsNet · CAMO | Fmaxβ: 0.634 -> 0.661 | Helped |
| add the Class-Head (ablation) | SdalsNet · CAMO | Fmaxβ: 0.218 -> 0.661 | Helped |
| use Frozen Training (ablation) | SdalsNet · CAMO | Fmaxβ: 0.581 -> 0.661 | Helped |
| include Lg in the loss (ablation) | SdalsNet · CAMO | Fmaxβ: 0.629 -> 0.661 | Helped |
| include Lf in the loss (ablation) | SdalsNet · CAMO | Fmaxβ: 0.539 -> 0.661 | Helped |
| include Lb in the loss (ablation) | SdalsNet · CAMO | Fmaxβ: 0.629 -> 0.661 | Helped |
| include Lp in the loss (ablation) | SdalsNet · CAMO | Fmaxβ: 0.636 -> 0.661 | Helped |
| use SdalsNet instead of UCOD-DA | UCOD · CAMO | Sα: 0.701 -> 0.696 | Hurt |
| use SdalsNet instead of UCOD-DA | UCOD · NC4K | Sα: 0.755 -> 0.738 | Hurt |
| use SdalsNet instead of UCOD-DA | UCOD · COD10K | Fmaxβ: 0.548 -> 0.572 | Helped |
Surprising findings
- Removing the Attention Shift Learning losses drops performance drastically (CAMO S-alpha 0.300), showing the shift objectives are essential.
Limitations
- Relies on DINO-pretrained ViT-Small and hand-designed attention localization and denoising heuristics.
- Evaluated only on COD benchmarks; the authors note generalization to other detection tasks is future work.
Experimental setup
Datasets: COD10K-Tr CAMO-Tr
Benchmarks: CHAMELEON CAMO COD10K NC4K
Metrics: S-measure E-measure (max) F-measure (max) MAE
Baselines: TokenCut SelfMask SpectralSeg A2S-v2 FOUND UCOD-DA DINO
Builds on: DINO FOUND UCOD-DA
Closest papers by meaning
- Camouflaged Object Detection (CVPR 2020 · similarity 0.89)
- Segment, Magnify and Reiterate: Detecting Camouflaged Objects the Hard Way (CVPR 2022 · similarity 0.88)
- Zoom in and Out: A Mixed-Scale Triplet Network for Camouflaged Object Detection (CVPR 2022 · similarity 0.88)
- Shift the Lens: Environment-Aware Unsupervised Camouflaged Object Detection (CVPR 2025 · similarity 0.88)
- UCOD-DPL: Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning (CVPR 2025 · similarity 0.87)
- Endow SAM with Keen Eyes: Temporal-spatial Prompt Learning for Video Camouflaged Object Detection (CVPR 2024 · similarity 0.86)
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.