EReCu: Pseudo-label Evolution Fusion and Refinement with Multi-Cue Learning for Unsupervised Camouflage Detection

Conference: CVPR 2026
arXiv: 2603.11521
Code: GitHub
Area: Unsupervised camouflaged object detection / Image segmentation
Keywords: unsupervised camouflaged object detection, pseudo-label evolution, multi-cue perception, teacher-student, spectral attention fusion

TL;DR

Ours proposes EReCu, a unified framework that utilizes Multi-cue Native Perception (MNP) to extract texture and semantic priors from the DINO teacher-student architecture. These priors guide Pseudo-label Evolution Fusion (PEF) and Local Pseudo-label Refinement (LPR) to recover boundary details. This work represents the first unification of pseudo-label guidance and feature learning paradigms in UCOD, achieving SOTA across four COD datasets.

Background & Motivation

Background: Camouflaged Object Detection (COD) is highly challenging due to the high similarity between targets and backgrounds. Fully supervised methods rely on expensive pixel-level annotations, which limits dataset scale and ecological diversity. Current Unsupervised COD (UCOD) follows two paradigms: pseudo-label guidance and feature learning.

Limitations of Prior Work:

1. Pseudo-label guidance paradigms (e.g., UCOS-DA, UCOD-DPL) rely excessively on high-dimensional embeddings while ignoring native image cues, leading to boundary overflow and semantic drift.
2. Feature learning paradigms (e.g., SdalsNet, EASE) lack explicit pseudo-label supervision, resulting in blurred boundaries and lost details.
3. Both paradigms have fatal flaws and have not yet been unified—semantic reliability and texture fidelity are optimized in isolation.

Key Challenge: Pseudo-label guidance addresses "where" but with inaccurate boundaries, while feature learning addresses "what it looks like" but with blurry localization—the two are complementary, yet existing methods cannot utilize both simultaneously.

Goal: Construct a unified UCOD framework where pseudo-label reliability and feature fidelity synergistically evolve through a mutual feedback loop.

Key Insight: Extract multi-cue native perception (texture + semantics) from original images to simultaneously constrain pseudo-label semantic evolution and local detail refinement.

Core Idea: Drive both global evolution and local refinement of pseudo-labels using native image cues to achieve semantic-perceptual co-evolution.

Method

Overall Architecture

EReCu aims to bridge the two UCOD routes—"pseudo-label guidance" (good at localization but blurry boundaries) and "feature learning" (good details but unstable localization)—into a closed loop of mutual data feeding. The system is built on a DINO teacher-student architecture: the Teacher produces a stable version via EMA (momentum 0.99), while the Student refines the segmentation mask iteratively.

When an unlabeled image enters, MNP extracts texture and semantic native cues from original pixels rather than high-dimensional embeddings and calculates a quality score \(S_{\text{mc}}\) to measure mask accuracy. These cues and scores drive two paths: PEF for global pseudo-label evolution and fusion, and LPR for boundary detail refinement. Inside PEF, EPL performs denoising through iterative student-teacher alignment, and STAF applies spectral fusion to multi-layer attention to suppress noise. LPR selects clean, focused attention heads from the Teacher to generate local pseudo-labels for boundary recovery. MNP constrains both paths, ensuring semantic reliability and texture fidelity rise together.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}}%%
flowchart TD
    IN["Unlabeled Image"] --> TS["DINO Teacher-Student Architecture<br/>Teacher uses EMA(0.99) to smooth Student"]
    TS --> MNP["Multi-Cue Native Perception (MNP)<br/>Texture(LBP+DoG) + Semantics(ResNet-18) → F_MNP<br/>Three-region patch sampling for quality score S_mc"]
    MNP -->|"F_MNP · S_mc Constraint"| PEF
    MNP -->|"F_MNP · S_mc Guidance"| LPR
    subgraph PEF["Pseudo-label Evolution Fusion (Global Pseudo-label)"]
        direction TB
        EPL["Evolutionary Pseudo-label Learning<br/>DSC for details + Teacher-Student mask alignment"]
        STAF["Spectral Tensor Attention Fusion<br/>Multi-layer Tucker + Truncated SVD denoising"]
        EPL --> STAF
    end
    PEF -->|"Global Pseudo-label M_fu"| LPR
    subgraph LPR["Local Pseudo-label Refinement (Boundary Details)"]
        direction TB
        TAS["Target-Aware Attention Selection<br/>Filter heads by entropy + cue consistency"]
        LPG["Local Pseudo-label Generation<br/>Adaptive thresholding → Dice + CE for boundaries"]
        TAS --> LPG
    end
    LPR --> OUT["Refined Camouflage Mask"]

Key Designs

1. Multi-Cue Native Perception (MNP): Finding flaws in camouflage via original textures

A common issue in prior work is focusing only on high-dimensional backbone embeddings, which often smooth out the subtle differences between target and background, leading to boundary overflow. MNP returns to the original image: it extracts low-level texture features \(F_{\text{text}}\) using LBP + DoG and mid-level semantic features \(F_{\text{sem}}\) using a frozen ResNet-18, concatenating them into \(F_{\text{MNP}} = \mathcal{C}(F_{\text{text}}, F_{\text{sem}})\).

To turn this into a supervision signal, MNP partitions the image into internal \(R_i\), boundary \(R_s\), and external \(R_o\) regions based on the current mask. It randomly samples \(K\times K\) patches across these regions over \(N\) rounds to calculate the corrected cosine similarity and aggregate the quality score:

\[S_{\text{mc}} = (D_{\text{io}} + D_{\text{is}} + S_{\text{so}}) / 3\]

The constraint loss is defined as \(\mathcal{L}_{\text{MNP}} = 1 - S_{\text{mc}}\). The intuition is that even in perfect camouflage, subtle distinguishable texture differences exist between internal and external regions. Forcing low similarity between internal/external regions and mid-level transitions pushes the mask toward true boundaries. Random patch sampling addresses irregular shapes that fixed grids cannot capture.

2. Pseudo-label Evolution Fusion (PEF): Mutual error correction between shallow details and deep semantics

Pseudo-labels from a single layer often lack either semantics or detail. PEF uses EPL and STAF to solve this. EPL (Evolutionary Pseudo-label Learning) passes Student shallow features through Depthwise Separable Convolution (DSC) to recover spatial details (\(M_s^{\text{dsc}}\)), then iteratively aligns \(M_s^{\text{dsc}}\) with pseudo-masks \(M_s^p\) and \(M_t^p\) from both branches while applying multi-cue constraints:

\[M_s^{\text{dsc}(r+1)} = \arg\min\big[\mathcal{L}_D(M_s^{\text{dsc}}, M_s^p) + \mathcal{L}_D(M_s^{\text{dsc}}, M_t^p) + \mathcal{L}_{\text{MNP}}\big]\]

STAF (Spectral Tensor Attention Fusion) addresses noisy single-layer attention maps. It stacks Student attention maps from three levels into a third-order tensor \(\mathcal{T}_s \in \mathbb{R}^{3 \times C \times HW}\), applies Tucker decomposition followed by truncated SVD to retain only the top \(t\) principal spectral components, resulting in a low-rank approximation \(A_s^{\text{fu}} = P_t \Sigma_t Q_t^\top\). This naturally discards high-frequency noise while preserving shared semantics and structure, with a complexity of only \(\mathcal{O}(r^2 d)\).

3. Local Pseudo-label Refinement (LPR): Using attention head diversity for boundary recovery

While global pseudo-labels locate the target center, they often miss boundary details. LPR leverages the spatial diversity of different Teacher attention heads. TAS (Target-Aware Attention Selection) calculates a focus entropy \(E_k\) and retains heads that are both clean (\(E_k < \tau_e\)) and consistent with native cues (\(S_{\text{mc}}(\hat{A}_k, F_{\text{MNP}}) > \tau_s\)). Thresholds are learnable.

LPG (Local Pseudo-label Generation) applies an adaptive threshold \(\tau_k = \mu_{A_k} + \alpha \cdot \sigma_{A_k}\) to selected heads to extract high-confidence regions. These form local pseudo-labels \(P_k\), which guide the refinement of the global prediction \(M_s^{\text{fu}}\) using Dice + CE losses.

Loss & Training

Total Loss = EPL Dice loss (aligning student DSC mask with student/teacher pseudo-masks) + \(\mathcal{L}_{\text{MNP}}\) (multi-cue constraint) + LPR Dice+CE loss (aligning fused prediction with local pseudo-labels). Training: 25 epochs, batch size 32, AdamW + Cosine Annealing, AMP. Backbone: DINO-ViT-S/8. Datasets: CAMO-Train (1000) + COD10K-Train (3040), unlabeled.

Key Experimental Results

Main Results

Comparison with UCOD Methods (4 COD Datasets)

Method	Type	CHAMELEON \(S_m\)↑	CAMO \(S_m\)↑	COD10K \(S_m\)↑	NC4K \(S_m\)↑
FOUND	UOS	.7161	.6913	.6783	.7459
UCOS-DA	UCOD	.6715	.6581	.6334	.7189
UCOD-DPL	UCOD	.7287	.7013	.7090	.7538
SdalsNet	UCOD	.7236	.6971	.6967	.7386
EReCu	UCOD	.7321	.7027	.7221	.7583

Ablation Study

Module Combinations (CAMO / COD10K \(S_m\)↑)

MNP	EPL	STAF	LPR	CAMO	COD10K
✓	✓	✓	✓	.7027	.7221
✗	✓	✓	✓	.6887	.7111
✓	✗	✗	✓	.6758	.7038
✓	✓	✓	✗	.6895	.7109
✗	✗	✗	✗	.6376	.6400

Key Findings

• The full module combination achieves UCOD SOTA across all major metrics on four datasets.
• PEF (including EPL+STAF) provides the largest contribution: removing it drops CAMO \(S_m\) by 2.69% (.7027→.6758).
• The MNP + EPL combination yields the greatest complementary gain, validating the role of native cues in guiding pseudo-label evolution.
• Single or double module performance is significantly lower than the full combination, confirming strong complementarity.
• DINO baseline (no modules): CAMO \(S_m = .6376\); our full model improves this by +.0651.

Highlights & Insights

• Unifies the two UCOD paradigms into a synergistic evolution framework, which is conceptually robust.
• The \(S_{\text{mc}}\) quality metric designed via the three-region (internal/boundary/external) patch sampling is ingenious and can be reused for mask quality estimation in other unsupervised tasks.
• STAF provides a lightweight and elegant solution (\(\mathcal{O}(r^2d)\)) for multi-scale feature aggregation using Tucker decomposition and SVD.
• The dual-condition selection in TAS (entropy + consistency) shows strong generalization capability.

Limitations & Future Work

• Gains on certain datasets/metrics are marginal (e.g., +.0014 for CAMO \(S_m\)), and performance on COD10K MAE is comparable to UCOD-DPL.
• Validated only on DINO-ViT-S/8; larger backbones like DINOv2 remain unexplored.
• Texture descriptors in MNP (LBP, DoG) are hand-crafted; learning-based substitutes could be explored.
• The training overhead of multi-loss, Tucker/SVD, and EMA is non-trivial.
• Handling of multi-instance camouflage scenarios was not discussed.

Related Work & Insights

• vs UCOD-DPL: Both use teacher-student dynamic pseudo-labels, but UCOD-DPL ignores native cues leading to overflow. EReCu introduces MNP for guidance and STAF for superior aggregation.
• vs SdalsNet: SdalsNet lacks pseudo-label supervision leading to blurry details. ERECu bridges both advantages.
• vs FOUND: FOUND uses a background-first paradigm which fails at fine-grained boundaries in high-similarity camouflage scenarios.
• Insights: The \(S_{\text{mc}}\) metric can be applied in active learning to estimate mask quality for unlabeled samples. The pseudo-label evolution + native cue guidance paradigm is transferable to unsupervised saliency or medical image segmentation.

Rating

• Novelty: ⭐⭐⭐⭐ Strong unification idea; individual module designs are innovative.
• Experimental Thoroughness: ⭐⭐⭐⭐ 4 datasets plus comprehensive ablation and visualization.
• Writing Quality: ⭐⭐⭐⭐ Clear framework diagram, complete formulas, and logical flow.
• Value: ⭐⭐⭐⭐ SOTA in UCOD with open-source code; \(S_{\text{mc}}\) and STAF are reusable components.

Related Papers

• [ECCV 2024] Learning Camouflaged Object Detection from Noisy Pseudo Label
• [CVPR 2026] Uncertainty-Aware Modality Fusion for Unaligned RGB-T Salient Object Detection
• [CVPR 2026] Unsupervised Multi-Scale Segmentation of 3D Subcellular World with Stable Diffusion Foundation Model
• [CVPR 2026] Love Me, Love My Label: Rethinking the Role of Labels in Prompt Retrieval for Visual In-Context Learning
• [NeurIPS 2025] Towards Robust Pseudo-Label Learning in Semantic Segmentation: An Encoding Perspective