Research map / Detection and segmentation / CNN features
RUN: Reversible Unfolding Network for Concealed Object 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 RUN, a deep unfolding network that unrolls a proximal-gradient solution of a foreground-background separation model with a residual sparsity constraint, applying reversible mask-level (SOFS) and RGB-level (ROBE) modules per stage for concealed object segmentation.
Problem
Concealed object segmentation must separate objects visually blended into backgrounds; existing reversible foreground-background strategies operate only on the mask domain, leaving RGB-domain cues underexplored.
Approach
Formulate COS as foreground-background separation with an extra residual sparsity constraint, unfold its proximal-gradient iterations into a multi-stage network, and pair mask-level refinement (SOFS with Reversible State Space) with an RGB reconstruction module (ROBE) that targets distortion-prone regions where foreground and background estimates conflict.
Key results
- On camouflaged object detection with ResNet50 in the common setting, RUN reports MAE 0.027 and S-measure 0.895 on CHAMELEON, and on COD10K MAE 0.030.
- Ablation on COD10K: RUN with 4 stages reaches MAE 0.030 and S-alpha 0.827, versus 0.033/0.803 with a single stage (K=1).
- RUN with a stronger backbone (PVT V2) is reported as achieving top results across polyp, tubular, and transparent object benchmarks, e.g., 0.822 mDice on CVC-ColonDB and 0.952 F-beta on GDD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use RUN instead of FocusDiff | COD segmentation with ResNet50 · COD10K | F_beta: 0.730 -> 0.747 | Helped |
| replace concealed image input with deep features (ablation) | RUN SOFS · COD10K | F_beta: 0.617 -> 0.747 | Helped |
| add RSS (ablation) | RUN SOFS · COD10K | F_beta: 0.710 -> 0.747 | Helped |
| add VSS (ablation) | RUN SOFS · COD10K | F_beta: 0.740 -> 0.747 | Helped |
| add prior mask M_hat_k (ablation) | RUN SOFS · COD10K | F_beta: 0.728 -> 0.747 | Helped |
| add prior background B_(k-1) (ablation) | RUN SOFS · COD10K | F_beta: 0.735 -> 0.747 | Helped |
| add auxiliary edge output (ablation) | RUN SOFS · COD10K | F_beta: 0.733 -> 0.747 | Helped |
| replace simple reconstruction network with CNN-based B1 (ablation) | RUN ROBE · COD10K | F_beta: 0.747 -> 0.746 | No clear effect |
| replace simple reconstruction network with Transformer-based B2 (ablation) | RUN ROBE · COD10K | F_beta: 0.747 -> 0.749 | No clear effect |
| add reconstructed output C_hat_k (ablation) | RUN ROBE · COD10K | F_beta: 0.736 -> 0.747 | Helped |
| make originally fixed parameters learnable (ablation) | RUN · COD10K | F_beta: 0.731 -> 0.747 | Helped |
| increase stage count from 1 to 4 (ablation) | RUN · COD10K | F_beta: 0.715 -> 0.747 | Helped |
Surprising findings
- Replacing the U-shaped reconstruction network with larger CNN or Transformer networks gave no significant gains, suggesting a simple network suffices.
- Many ablation variants that differ in the uncertainty-range mapping produce only small differences, indicating limited sensitivity to those design choices.
Limitations
- Performance is unstable in degraded scenarios such as haze, as with other advanced methods.
- Unfolding multi-stage design with large low-level algorithms adds computational and storage cost.
Experimental setup
Datasets: CAMO COD10K CHAMELEON NC4K CVC-ColonDB ETIS DRIVE CORN GDD GSD CDS2K DUT-OMRON
Benchmarks: CHAMELEON CAMO COD10K NC4K CVC-ColonDB ETIS DRIVE CORN GDD GSD CDS2K DUT-OMRON
Metrics: MAE F-measure E-measure S-measure mDice mIoU AUC sensitivity
Baselines: SINet FEDER FSEL FocusDiff FGANet CamoFocus HitNet LSSNet
Builds on: FEDER SegMaR CoRUN Visual State Space (VMamba)
Closest papers by meaning
- Weakly-Supervised Concealed Object Segmentation with SAM-based Pseudo Labeling and Multi-scale Feature Grouping (NeurIPS 2023 · similarity 0.86)
- Shift the Lens: Environment-Aware Unsupervised Camouflaged Object Detection (CVPR 2025 · similarity 0.84)
- Any3DIS: Class-Agnostic 3D Instance Segmentation by 2D Mask Tracking (CVPR 2025 · similarity 0.84)
- Alligat0R: Pre-Training through Covisibility Segmentation for Relative Camera Pose Regression (NeurIPS 2025 · similarity 0.84)
- unMORE: Unsupervised Multi-Object Segmentation via Center-Boundary Reasoning (ICML 2025 · similarity 0.84)
- FOCUS: Towards Universal Foreground Segmentation (AAAI 2025 · similarity 0.84)
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.