Skip to content

PASR: Pattern-Aware Scene-Conditioned Reasoning for Camouflaged Object Detection

Conference: ECCV 2026
Paper: ECCV 2026 Poster 5521
Area: Object Detection / Segmentation
Keywords: Camouflaged Object Detection, Scene-Conditioned Reasoning, Prototype Learning, Unsupervised Inference, Background Deviation Modeling

TL;DR

PASR introduces an annotation-free, zero-training framework for camouflaged object detection that reframes the task from foreground discrimination into scene-conditioned pattern-deviation reasoning over a frozen feature space, significantly outperforming existing unsupervised/weakly supervised methods and closing the gap to fully supervised models.

Background & Motivation

Camouflaged Object Detection (COD) aims to identify concealed objects that visually blend into their surrounding natural environment. Through evolutionary adaptation, camouflaged organisms deliberately conform their surface patterns, textures, and color distributions to local habitat statistics. This mechanism causes extreme statistical overlap between foreground and background feature distributions, fundamentally breaking the classical assumption in generic object detection and semantic segmentation that foreground objects exhibit distinctive global saliency or category-level appearance consistency.

Existing COD methodologies generally follow two main paradigms. The predominant object-centric discriminative paradigm designs sophisticated multi-scale feature aggregators, boundary-aware decoders, or frequency decomposition modules to forcibly separate foreground from background. However, these heavily supervised architectures depend strictly on dense pixel annotations and suffer severe performance degradation when confronted with unseen biological species or shifted background distributions. On the other hand, recent reference-augmented methods (such as RISE and EASE) leverage external exemplar libraries or environmental prototypes to assist detection. Yet, they predominantly treat object appearances and background statistics as independent entities, overlooking the intrinsic truth that camouflage is an adaptive, structured perturbation conditioned on specific background contexts.

This paper's core insight is that a camouflaged target is not an isolated visual primitive, but rather a structured deviation constrained by background statistics; hence, evaluating object presence requires modeling relative pattern deviations under a scene-specific background condition. Core idea: reframe camouflaged object detection as background-conditioned pattern-deviation reasoning within a frozen representation space, constructing an offline annotation-free prototype library and executing a two-stage "global background anchoring - patch-level deviation reasoning - query self-refinement" pipeline without requiring any model retraining.

Method

Overall Architecture

PASR comprises two primary components: offline unsupervised prototype library construction and progressive scene-conditioned reasoning at inference time. The entire system operates inside the frozen embedding space of a DINOv2-ViT-L/14 backbone without back-propagation or fine-tuning. Offline, reference images harvested from web natural platforms undergo embedding clustering and Kernel Density Estimation (KDE) to yield decoupled foreground-background prototype pairs without human annotation. During online inference, a query image first undergoes coarse background subspace alignment (BG-anchor) to discard irrelevant contexts, followed by patch-level background-conditioned deviation prototype matching and KDE statistical thresholding. Finally, a query self-refinement step purifies cross-image domain noise to yield high-quality segmentation masks.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Query Image & Offline Prototype Library"] --> B["Stage 1: Coarse Background Subspace Alignment<br/>Construct Scene BG-anchor Candidate Set"]
    B --> C["Stage 2: Background-Conditioned Pattern Reasoning<br/>Patch-level Deviation Prototype Matching & KDE"]
    C --> D["Stage 3: Query Self-Reasoning Refinement<br/>Query-specific Contrastive Filtering"]
    D --> E["Final Camouflaged Object Mask<br/>(Optional SAM Geometric Refinement)"]

Key Designs

1. Unsupervised Prototype Library Construction: Annotation-Free Clustering Self-Reasoning To bypass the heavy reliance of external exemplar libraries on manual pixel annotations, PASR develops a self-reasoning prototype extraction mechanism based on unsupervised clustering. For an unannotated reference image \(R_t\), dense embeddings \(F_t \in \mathbb{R}^{H \times W \times d}\) are extracted using frozen DINOv2. Following k-means clustering in feature space, clusters exhibiting large spatial extent with centroids situated near the image geometric center are pruned (as they typically indicate central objects), leaving the remaining groups as candidate background clusters. The spatial similarity against these candidates yields an internal consistency response map: $\(S_t^{\text{self}}(x, y) = \max_k \frac{F_t(x, y) \cdot \tilde{P}_{t, k}^{\text{bg}}}{\|F_t(x, y)\| \|\tilde{P}_{t, k}^{\text{bg}}\|}\)$ Using Kernel Density Estimation (KDE) over \(S_t^{\text{self}}\), an adaptive threshold \(\tau_t\) isolates the coarse foreground mask. A frozen SAM instance is then prompted with this coarse mask for purely structural boundary regularization without semantic supervision. Weighted spatial pooling over the regularized regions produces decoupled prototype pairs \((P_t^{\text{fg}}, P_t^{\text{bg}})\). The whole procedure requires zero manual ground-truth labels.

2. Background-Conditioned Pattern Reasoning: Modeling Camouflage as Structured Deviation Searching across the entire reference library directly would inevitably introduce cross-scene contextual interference due to disparate environmental statistics. PASR first derives an image-level global query descriptor \(F_j^g\) via average pooling and evaluates cosine similarities against all library background prototypes \(P_t^{\text{bg}}\), discarding the bottom 50% divergent candidates to form a scene-aligned background anchor subspace \(\mathcal{R}_j\). At the patch level, for each spatial feature \(F_j(x, y)\), the top-\(K\) background-matched references \(\mathcal{T}_j(x, y)\) are retrieved, and their corresponding foreground vectors are aggregated into a query-adaptive object prototype \(\bar{P}_j^{\text{obj}}(x, y)\). To explicitly formulate camouflage as an environmental perturbation, a background-conditioned pattern prototype \(P_{jt}^{\text{pat}}\) is synthesized: $\(P_{jt}^{\text{pat}}(x, y) = P_t^{\text{bg}} - \alpha \bar{P}_j^{\text{obj}}(x, y), \quad t \in \mathcal{T}_j(x, y)\)$ where \(\alpha\) is a deviation scaling parameter (set to 0.4 by default). Evaluating cosine similarities across these deviation prototypes produces response values \(S_j(x, y)\), wherein true camouflaged regions exhibit prominent deviation troughs. An empirical KDE thresholding operation directly extracts the initial detection mask \(M_j^{(0)}\).

3. Query Self-Reasoning Refinement: Filtering Cross-Image Distributional Noise While cross-image retrieval establishes strong contextual priors, unavoidable variations in illumination, sensor calibration, and microscopic textures introduce cross-scene domain noise into \(M_j^{(0)}\). To filter these artifacts, an internal self-referential cycle is enacted directly on the query image. Using \(M_j^{(0)}\), query-specific foreground and background prototypes \(\hat{P}_j^{\text{fg}}\) and \(\hat{P}_j^{\text{bg}}\) are computed via masked spatial averaging over query features \(F_j\). Each patch is then classified by contrasting its relative cosine affinities: $\(\Delta_j(x, y) = \text{sim}(F_j(x, y), \hat{P}_j^{\text{bg}}) - \text{sim}(F_j(x, y), \hat{P}_j^{\text{fg}})\)$ A subsequent KDE separation on \(\Delta_j(x, y)\) produces the refined mask \(M_j^{(1)}\). This step enforces internal statistical consistency, effectively suppressing spurious activations caused by imperfect cross-scene reference matches.

Loss & Training

PASR operates in a completely training-free and backpropagation-free regime with all backbone parameters permanently frozen. On a single NVIDIA A40 GPU, core inference latency ranges between 71 and 76 ms per image (comprising ~51 ms for DINOv2 feature extraction, ~13 ms for conditioned reasoning, and ~10 ms for query self-refinement).

Key Experimental Results

Main Results

On four widely acknowledged COD benchmarks (CHAMELEON, CAMO, COD10K, and NC4K), PASR is evaluated against leading unsupervised methods under a unified DINOv2-ViT-L/14 backbone and standardized 476×476 input resolution without test-time augmentation:

Dataset Metric Ours (PASR) Prev. SOTA (EASE [CVPR 2025]) Gain
CHAMELEON \(S_\alpha \uparrow\) 0.832 0.819 +0.013
CHAMELEON \(E_\phi \uparrow\) 0.924 0.899 +0.025
CHAMELEON \(F_\beta^w \uparrow\) 0.768 0.741 +0.027
CHAMELEON \(\text{MAE} \downarrow\) 0.041 0.044 -0.003
CAMO \(S_\alpha \uparrow\) 0.755 0.749 +0.006
CAMO \(E_\phi \uparrow\) 0.833 0.831 +0.002
CAMO \(F_\beta^w \uparrow\) 0.693 0.684 +0.009
COD10K \(S_\alpha \uparrow\) 0.774 0.773 +0.001
COD10K \(E_\phi \uparrow\) 0.868 0.866 +0.002
COD10K \(F_\beta^w \uparrow\) 0.664 0.656 +0.008
NC4K \(S_\alpha \uparrow\) 0.817 0.800 +0.017
NC4K \(E_\phi \uparrow\) 0.902 0.884 +0.018
NC4K \(F_\beta^w \uparrow\) 0.760 0.735 +0.025
NC4K \(\text{MAE} \downarrow\) 0.051 0.056 -0.005

When equipped with downstream SAM prompt-based boundary refinement, PASR scores \(S_\alpha=0.869, E_\phi=0.937, F_\beta^w=0.846\) on CHAMELEON and \(S_\alpha=0.863, E_\phi=0.917, F_\beta^w=0.840\) on NC4K, comfortably outperforming promptable competitors GenSAM and ProMaC while closing the performance gap to fully supervised models like SINet-v2 (\(S_\alpha=0.847\) on NC4K).

Ablation Study

Ablation experiments across benchmarks detail the individual contribution of each architectural component (evaluated by \(S_\alpha / F_\beta^w\)):

Config CHAM. (\(S_\alpha / F_\beta^w\)) CAMO (\(S_\alpha / F_\beta^w\)) COD10K (\(S_\alpha / F_\beta^w\)) NC4K (\(S_\alpha / F_\beta^w\)) Note
Full PASR 0.832 / 0.768 0.755 / 0.693 0.774 / 0.664 0.817 / 0.760 full model
\(\alpha = 0\) (pure background matching) 0.819 / 0.745 0.741 / 0.674 0.767 / 0.654 0.798 / 0.734 severe drop; loses conditioned deviation discriminability
w/o adaptive object prototype 0.825 / 0.755 0.753 / 0.691 0.773 / 0.662 0.813 / 0.755 degradation from unadapted foreground signals
w/o query self-refinement 0.826 / 0.756 0.751 / 0.687 0.769 / 0.656 0.808 / 0.748 loss of intra-image filtering increases cross-scene noise
w/o BG-anchor (full-library search) 0.829 / 0.764 0.752 / 0.689 0.776 / 0.664 0.816 / 0.759 context aliasing from diverse background statistics
random BG-anchor 0.820 / 0.744 0.750 / 0.687 0.770 / 0.656 0.815 / 0.757 distorted background conditions harm inference
w/o offline SAM refinement 0.829 / 0.764 0.754 / 0.692 0.774 / 0.662 0.812 / 0.752 library built purely from coarse clusters remains robust

Key Findings

  • Deviation formulation is the core performance driver: Disabling the deviation coefficient (\(\alpha=0\)) degrades the model into naive background matching, causing \(F_\beta^w\) on NC4K to plummet from 0.760 to 0.734. This underscores that explicitly modeling foreground as a conditioned perturbation over background statistics is vital for camouflage reasoning.
  • Remarkable efficiency with minimal reference scaling: Prior methods rely on massive candidate pools (e.g., EASE constructs 16,000 environment prototypes and retrieves 1,024 candidates; RISE retrieves 512). In contrast, PASR achieves superior results using only 315 natural web images and retrieving top \(K=32\) candidates.
  • Independence from precise mask supervision: Omitting SAM geometric regularization during offline library construction causes negligible performance loss (\(S_\alpha\) remains at 0.812 on NC4K vs. 0.817 for the full model), verifying that accurate prototype deviation modeling depends on proper statistical conditioning rather than pristine boundary annotations.

Highlights & Insights

  • Paradigm shift (deviation over saliency): Transcends the traditional discriminative trap of searching for distinctive foreground signatures by formulating camouflage detection as identifying structured perturbations on a background manifold.
  • Extremely lean footprint and zero annotation: By leveraging a simple algebraic formulation (\(P_t^{\text{bg}} - \alpha \bar{P}_j^{\text{obj}}\)) and unsupervised clustering, PASR achieves SOTA accuracy using just 315 web reference images, 32 retrieved candidates, and a rapid 75 ms inference budget.
  • Generalizable inductive pattern: The three-step inference workflow—global manifold anchoring, local conditioned perturbation reasoning, and intra-query self-filtering—offers a compelling blueprint for other ill-posed visual discovery problems, such as surface anomaly detection and unannotated lesion discovery.

Limitations & Future Work

  • Confusion under severe shadows and heavy occlusions: Dense canopy shadows or severe physical occlusions can exhibit statistical deviations comparable to camouflaged organisms, occasionally triggering false-positive detections.
  • Uniform patch resolution vs. micro-scale targets: The fixed patch tokenization in DINOv2-ViT-L/14 can smooth out tiny camouflaged insects or millimeter-level textural variations during spatial pooling. Incorporating multi-scale adaptive patch pooling or dynamic windowing represents a promising future avenue.
  • vs EASE (CVPR 2025): While EASE constructs environmental libraries to analyze background deviations independently, PASR establishes an explicit background-conditioned coupling (\(P_t^{\text{bg}} - \alpha \bar{P}_j^{\text{obj}}\)), slashing both library size (315 vs. 16,000) and retrieval overhead (32 vs. 1,024) while achieving higher accuracy.
  • vs RISE (ICCV 2025): RISE relies on heavy self-augmented retrieval without modeling the conditional dependency between background and object appearance. PASR couples the two representations, unlocking superior structural fidelity on intricate benchmarks like NC4K.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ [Reformulating camouflage detection as scene-conditioned deviation reasoning fundamentally corrects the flawed assumption of independent foreground saliency]
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Extensive multi-benchmark validation covering unsupervised, weakly supervised, and prompt-based regimes alongside comprehensive ablations]
  • Writing Quality: ⭐⭐⭐⭐⭐ [Clean formulation, rigorous logical motivation, and excellent consistency between mathematical modeling and empirical outcomes]
  • Value: ⭐⭐⭐⭐⭐ [Presents an efficient, annotation-free, and deployable paradigm for high-precision camouflaged object segmentation]