Research map / Detection and segmentation / Weakly supervised detection and segmentation
Beyond Weak Supervision: MLLMs-Guided Graded Knowledge Distillation for Unsupervised Camouflaged Object Detection
Huafeng Chen, Chenguang Zhu, Yueming Lyu, Caifeng Shan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 CA-CoT prompting to MLLM+SAM teacher (ablation) | MLLM + SAM teacher · COD10K | MAE: 0.232 -> 0.085 | Improved |
| add Graded Mask Evaluator to filter low-quality masks (ablation) | UCOD-MKD with CA-CoT · COD10K | MAE: 0.085 -> 0.041 | Improved |
| add GKD distillation (ablation) | UCOD-MKD · COD10K | MAE: 0.041 -> 0.031 | Improved |
| add each CA-CoT reasoning step (STEP1 to STEP4) (ablation) | CA-CoT · COD10K | MAE: 0.232 -> 0.085 | Improved |
| add SIM-based mask grading in GME (ablation) | GME baseline · COD10K | MAE: 0.085 -> 0.066 | Improved |
| add both FGC and ETC to SIM (ablation) | GME · COD10K | MAE: 0.066 -> 0.041 | Improved |
| use SSIM together with IoU in SIM (ablation) | GME SIM · COD10K | MAE: 0.071 -> 0.066 | Improved |
| add pixel-level enhancement (S and M) (ablation) | GKD · COD10K | MAE: 0.041 -> 0.038 | Improved |
| add L1 and MSE losses for high-quality samples (ablation) | GKD image-level enhancement · COD10K | MAE: 0.041 -> 0.035 | Improved |
| add low-quality samples to training (ablation) | COD distillation · CAMO+COD10K training | E-measure | Hurt |
| use fewer high-quality samples instead of more random samples (ablation) | COD distillation · E-measure vs training samples | E-measure | Improved |
| replace DINO backbone dependence with MLLM+SAM supervision | UCOD-MKD vs UCOD-DPL · COD10K | MAE: 0.059 -> 0.031 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- UCOD-DPL — UCOD-DPL: Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning (CVPR 2025) · better on 6 of 6 shared results
- SAM-COD — SAM-COD: SAM-guided Unified Framework for Weakly-Supervised Camouflaged Object Detection (ECCV 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SAM-COD: SAM-guided Unified Framework for Weakly-Supervised Camouflaged Object Detection (ECCV 2024 · similarity 0.92)
- Learning Camouflaged Object Detection from Noisy Pseudo Label (ECCV 2024 · similarity 0.91)
- Just a Hint: Point-Supervised Camouflaged Object Detection (ECCV 2024 · similarity 0.89)
- Weakly-Supervised Camouflaged Object Detection with Scribble Annotations (AAAI 2023 · similarity 0.89)
- UCOD-DPL: Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning (CVPR 2025 · similarity 0.89)
- PASR: Pattern-Aware Scene-Conditioned Reasoning for Camouflaged Object Detection (ECCV 2026 · 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.
