Context-measure: Contextualizing Metric for Camouflage

TL;DR

This work tackles the context-dependence of camouflage evaluation in COS by criticizing context-independent metrics and introducing Context-measure, a pixel-correlation-based framework that models forward inference and reverse deduction via a reciprocal perceptual loop.A pixel-level camouflage degree is learned in LAB space to contextualize the measure for camouflaged scenarios, producing a camouflage-aware variant $C_{eta}^{\omega}$ that better aligns with human perception.The authors validate their approach on three COS benchmarks and a new human-ranked dataset (CamoHR), showing superior consistency with human judgments and greater sensitivity to semantic and boundary variations than existing metrics.The work provides a practical benchmark and codebase for COS evaluation and suggests broader applicability of context-aware metrics to camouflage and related vision tasks.

Abstract

Camouflage is primarily context-dependent yet current metrics for camouflaged scenarios overlook this critical factor. Instead, these metrics are originally designed for evaluating general or salient objects, with an inherent assumption of uncorrelated spatial context. In this paper, we propose a new contextualized evaluation paradigm, Context-measure, built upon a probabilistic pixel-aware correlation framework. By incorporating spatial dependencies and pixel-wise camouflage quantification, our measure better aligns with human perception. Extensive experiments across three challenging camouflaged object segmentation datasets show that Context-measure delivers more reliability than existing context-independent metrics. Our measure can provide a foundational evaluation benchmark for various computer vision applications involving camouflaged patterns, such as agricultural, industrial, and medical scenarios. Code is available at https://github.com/pursuitxi/Context-measure.

Figures (10)

• Figure 1: Camouflage vs. Saliency. The red butterfly rests within a cluster of red maple leaves, blending with its background, while the blue butterfly perches amid yellow maple leaves, standing out clearly. The ground-truth mask is shown in Appendix.
• Figure 1: Camouflage vs. Saliency. The red butterfly rests within a cluster of red maple leaves, blending with its background, while the blue butterfly perches amid yellow maple leaves, standing out clearly.
• Figure 2: Flaws of context-independent metrics in Camouflage. Compared with two widely-used saliency metrics ($E_{\phi}$ & $S_{\alpha}$), our Context-measure ($C_{\beta}^{\omega}$) more effectively distinguishes masks in camouflaged scenarios, aligning more closely with human perception.
• Figure 3: Our general Context-measure framework. We formulate the foreground evaluation problem into a reciprocal perceptual loop, which alternates between two processes: forward inference$F(Y|X)$ and reverse deduction$R(X|Y)$. Intuitively, the former understands what the prediction conveys about reality, and the latter verifies how reality is reflected in the prediction. Further details are provided in §\ref{['sec:cmeasure']}.
• Figure 4: Visualization of the Camouflage Degree. Warmer colors (toward red) indicate a higher degree of camouflage, and vice versa.