Discover, Segment, and Select: A Progressive Mechanism for Zero-shot Camouflaged Object Segmentation

Yilong Yang; Jianxin Tian; Shengchuan Zhang; Liujuan Cao

cs.CVFeb 23, 2026

Discover, Segment, and Select: A Progressive Mechanism for Zero-shot Camouflaged Object Segmentation

Authors: Yilong Yang, Jianxin Tian, Shengchuan Zhang, Liujuan Cao

Organizations: Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, 361005, P.R. China.

Abstract

Current zero-shot Camouflaged Object Segmentation methods typically employ a two-stage pipeline (discover-then-segment): using MLLMs to obtain visual prompts, followed by SAM segmentation. However, relying solely on MLLMs for camouflaged object discovery often leads to inaccurate localization, false positives, and missed detections. To address these issues, we propose the Discover-Segment-Select (DSS) mechanism, a progressive framework designed to refine segmentation step by step. The proposed method contains a Feature-coherent Object Discovery (FOD) module that leverages visual features to generate diverse object proposals, a segmentation module that refines these proposals through SAM segmentation, and a Semantic-driven Mask Selection (SMS) module that employs MLLMs to evaluate and select the optimal segmentation mask from multiple candidates. Without requiring any training or supervision, DSS achieves state-of-the-art performance on multiple COS benchmarks, especially in multiple-instance scenes.

Figures & tables

Figure 1: Comparison between the proposed DSS framework and prior zero-shot COS methods.

Figure 2: The overall pipeline of the Discovery-Segment-Selection Framework. The framework operates in three stages: (a) Feature-guided Object Discovery (FOD): the Part Composition (PC) module refines low-resolution clustering masks, followed by a Similarity based Box Generation (SBG) module to produce prompt boxes; (b) The Segment Anything Model generates segmentation mask based on the input prompts. (c) Semantic-driven Mask Selection (SMS): an MLLM evaluates all candidate masks and selects the final segmentation.

Figure 3: Visual tracking of the iterative refinement process. Left to right: the original image, Leiden clustering map, and binary maps from initial clustering $\mathbf{Y}^{(0)}$ to the final refined map $\mathbf{Y}^{(5)}$ . Here we select one cluster from the clustering results for demonstration. The number of iterations may vary for different clusters until convergence.

Figure 4: Visual comparison between bboxes generated from different strategies. In each 2 $\times$ 2 grid, Top left : bboxes from the ground truth mask. Top right : bboxes from leiden clustered mask. Bottom left : bboxes from PC refined mask. Bottom right : bboxes from self-similarity map.

HP	Value	Remark
$r$	0.5	Leiden clustering resolution.
$\epsilon$	1.0	The energy change threshold.
$\tau$	0.95	Correlation threshold for de-duplication.
$K$	5	Top-K masks for MLLM mask selection.

Table 1: Hyperparameter setting.

Figure 5: We present a visual comparison of our method against existing alternatives under challenging conditions. The red and black bounding boxes correspond to prompts generated by our approach and QWen, respectively. To aid interpretation, the third column visualizes the similarity maps from which our bounding boxes are derived.

Methods	Venue	Sup	CAMO(250)				NC4K(4121)				COD10K(2016)				CHAMELEON(87)
Methods	Venue	Sup	$F_{\beta}^{w}\uparrow$	$\mathcal{M}\downarrow$	$S_{\alpha}\uparrow$	$E_{\phi}\uparrow$	$F_{\beta}^{w}\uparrow$	$\mathcal{M}\downarrow$	$S_{\alpha}\uparrow$	$E_{\phi}\uparrow$	$F_{\beta}^{w}\uparrow$	$\mathcal{M}\downarrow$	$S_{\alpha}\uparrow$	$E_{\phi}\uparrow$	$F_{\beta}^{w}\uparrow$	$\mathcal{M}\downarrow$	$S_{\alpha}\uparrow$	$E_{\phi}\uparrow$
HitNet [ 16 ]	AAAI23	F	.809	.055	.849	.906	.834	.037	.875	.926	.806	.023	.871	.935	.897	.019	.921	.967
SENet [ 12 ]	TIP24	F	.847	.039	.888	.932	.843	.032	.889	.933	.780	.024	.865	.925	.878	.019	.918	.957
ZoomNeXt [ 30 ]	TPAMI24	F	.857	.041	.889	.945	.863	.028	.903	.951	.827	.018	.898	.956	.885	.018	.924	.975
CGCOD [ 45 ]	MM25	F	.864	.036	.896	.947	.869	.026	.904	.949	.824	.018	.890	.948	.902	.017	.931	.972
CAMF [ 19 ]	TPAMI25	F	.887	.041	.901	.933	.897	.033	.905	.933	.845	.025	.890	.923	.895	.028	.917	.944

Table 2: Quantitative comparison with state-of-the-art methods on four COD datasets “

\uparrow/\downarrow

” : the higher/lower the better. “-”: not available. “F”, “U” and “ZS” denote “Fully-supervised”, “Un-supervised” and “Zero-shot” method, repsectively. The best three results in zero-shot setting are highlighted in red , blue , and green , respectively. The best results in “F”, “U” settings are in bold.

Panel of Figure 6 — Figure 6: (a) Each pie chart shows the proportion of images containing different numbers of camouflaged objects. “3+” denotes images contain $(\geq 3)$ camouflaged objects. (b) and (c) Performance comparison of different camouflaged instance counts on NC4K and COD10K. The y-axis is $score=\frac{S_{\alpha}+E_{\phi}+F_{\beta}^{w}}{3}-\mathcal{M}$ , indicateing the overall performance.

	ProMaC	RDVP-MSD	DSS
Time (s)	130.49	18.05	41.96 (7.74, 3.63, 30.59)
GPU (GB)	32.85	31.51	17.90

Table 3: Comparison of computational efficiency, reported as average inference time per image (s). For our DSS framework, the time is decomposed into its three stages: FOD, mask generation, and SMS. Results are reported on the CHAMELEON dataset.

ID	Leiden	PC	VLOS	COD10K	CHAMELEON	CAMO
1	$\checkmark$	$\times$	$\times$	.778	.797	.689
2	$\times$	$\checkmark$	$\times$	.758	.814	.761
3	$\times$	$\times$	$\checkmark$	.850	.791	.708
4	$\checkmark$	$\checkmark$	$\times$	.820	.863	.790
5	$\checkmark$	$\times$	$\checkmark$	.886	.872	.801
6	$\times$	$\checkmark$	$\checkmark$	.881	.880	.810

Table 4: Ablation comparison of different inference settings. “Leiden” and “PC” denotes that the initial Leiden clustering map or the refined clustering map is used in the FOD module for extracting bboxes; “VLOS” denotes that the mask produced by QWen+SAM2 is included for mask selection.

Methods	NC4K 3+				COD10K 3+
Methods	$F_{\beta}^{w}\uparrow$	$\mathcal{M}\downarrow$	$S_{\alpha}\uparrow$	$E_{\phi}\uparrow$	$F_{\beta}^{w}\uparrow$	$\mathcal{M}\downarrow$	$S_{\alpha}\uparrow$	$E_{\phi}\uparrow$
Leiden+PC	.795	.058	.812	.897	.723	.034	.794	.881
Sim. Map	.818	.054	.839	.917	.745	.026	.821	.911

Table 5: Ablation on SBG. The segmentation performance based on bboxes generated from: (1) Clustered and refined masks. (Leiden+PC). (2). similarity maps (Sim. Map). “ 3+ ” denotes the subset of images containing three or more camouflaged objects.

Method	CAMO	NC4K	COD10K	CHAMELEON
VLOS	.708	.845	.850	.791
DSS	.737	.869	.871	.852
Ideal Seg.	.823	.895	.892	.888

Table 6: Ablation on the SMS module.

Method	CAMO	CHAMELEON	Time (s)
Heuristic Rules	68.80	71.05	1.14
Single-pass Selection	OOM	OOM	OOM
Top- $K$ Single-pass	40.24	36.84	14.25
Top- $K$ pairwise (random)	54.00	59.21	31.14
Top- $K$ pairwise (descending)	50.80	43.42	30.81
Top- $K$ pairwise (ascending)	74.40	73.68	30.59

Table 7: Comparison of different Mask Selection strategies. CAMO and CHAMELEON columnes measure selection accuracy (%), while Time indicates average inference cost (seconds per image). “OOM” denotes out-of-memory errors.

Panel of Figure 7 — Figure 7: Prompt template for camouflaged object localization using MLLM.

Panel of Figure 8 — Figure 8: Prompt template for mask selection using MLLM.

Explore similar work

Phantom-Insight: Adaptive Multi-cue Fusion for Video Camouflaged Object Detection with Multimodal LLM
Sep 8, 2025Hua Zhang, Changjiang LuoVideo Object Segmentation Vision Foundation Model Adaptation
Institute of Information Engineering, Chinese Academy of Sciences · School of Cyber Security, University of Chinese Academy of Sciences
Learning Accurate Segmentation Purely from Self-Supervision
Feb 27, 2026Zuyao You, Zuxuan Wu, Yu-Gang JiangUnsupervised Image Segmentation Salient Object Detection
Fudan University
Language-Guided Structure-Aware Network for Camouflaged Object Detection
Mar 25, 2026Min ZhangCamouflaged Object Detection Boundary-Aware Segmentation
School of Artificial Intelligence, Chongqing University of Technology, Chongqing, 401120, China

Sep 8, 2025J/K move · Enter open · S save

Phantom-Insight: Adaptive Multi-cue Fusion for Video Camouflaged Object Detection with Multimodal LLM

Hua Zhang, Changjiang Luo

Institute of Information Engineering, Chinese Academy of Sciences · School of Cyber Security, University of Chinese Academy of Sciences

Video Object Segmentation Vision Foundation Model Adaptation Camouflaged Object Detection

Open paper AskPDF arXiv

Video camouflaged object detection (VCOD) is challenging due to dynamic environments. Existing methods face two main issues: (1) SAM-based methods struggle to separate camouflaged object edges due to model freezing, and (2) MLLM-based methods suffer from poor object separability as large language models merge foreground and background. To address these issues, we propose a novel VCOD method based on SAM and MLLM, called Phantom-Insight. To enhance the separability of object edge details, we represent video sequences with temporal and spatial clues and perform feature fusion via LLM to increase information density. Next, multiple cues are generated through the dynamic foreground visual token scoring module and the prompt network to adaptively guide and fine-tune the SAM model, enabling it to adapt to subtle textures. To enhance the separability of objects and background, we propose a decoupled foreground-background learning strategy. By generating foreground and background cues separately and performing decoupled training, the visual token can effectively integrate foreground and background information independently, enabling SAM to more accurately segment camouflaged objects in the video. Experiments on the MoCA-Mask dataset show that Phantom-Insight achieves state-of-the-art performance across various metrics. Additionally, its ability to detect unseen camouflaged objects on the CAD2016 dataset highlights its strong generalization ability.

Abstract

Explore similar work

Phantom-Insight: Adaptive Multi-cue Fusion for Video Camouflaged Object Detection with Multimodal LLM

Learning Accurate Segmentation Purely from Self-Supervision

Language-Guided Structure-Aware Network for Camouflaged Object Detection