Abstract: Aleatoric uncertainty captures noise within the observations. For camouflaged object detection, due to similar appearance of the camouflaged foreground and the background, it’s difficult to obtain highly accurate annotations, especially annotations around object boundaries. We argue that training directly with the "noisy" camouflage map may lead to a model of poor generalization ability. In this paper, we introduce an explicitly aleatoric uncertainty estimation technique to represent predictive uncertainty due to noisy labeling. Specifically, we present a confidence-aware camouflaged object detection (COD) framework using dynamic supervision to produce both an accurate camouflage map and a reliable "aleatoric uncertainty". Different from existing techniques that produce deterministic prediction following the point estimation pipeline, our framework formalises aleatoric uncertainty as probability distribution over model output and the input image. We claim that, once trained, our confidence estimation network can evaluate the pixel-wise accuracy of the prediction without relying on the ground truth camouflage map. Extensive results illustrate the superior performance of the proposed model in explaining the camouflage prediction.
1. Introduction
Camouflaged Object Detection (COD) involves identifying objects that are highly similar to their surroundings, such as animals with natural camouflage or human-made camouflaged targets. A significant challenge in this field is the inherent ambiguity of object boundaries. Because the foreground and background are visually indistinct, human annotators often struggle to provide precise ground truth labels. Standard deep learning approaches typically treat these noisy labels as absolute truth, which can lead to overconfident yet inaccurate predictions and poor generalization.
This research addresses the problem by explicitly modeling aleatoric uncertainty—the noise inherent in the data and labels. Instead of a purely deterministic output, the proposed framework estimates the confidence level for each predicted pixel, allowing the model to acknowledge and learn from its own uncertainty.
2. Proposed Method: Confidence-Aware COD Network (CANet)
The authors introduce the Confidence-Aware COD Network (CANet), a framework designed to jointly predict the camouflage map and its associated uncertainty.
2.1 Architecture
As shown in Figure 4, CANet consists of two primary components:
- CODNet (The Detector): This module utilizes a ResNet-50 backbone to extract features. It employs a Fusion Module to integrate high-level semantic information with low-level spatial details, producing an initial camouflage prediction.
- OCENet (The Uncertainty Estimator): A U-Net-style network that takes both the original image and the CODNet's prediction as input. Its goal is to generate a pixel-wise confidence map, representing the model's certainty about its own output.
2.2 Dynamic Supervision
A key innovation is the use of Dynamic Supervision. During the training phase, the error between the current prediction and the ground truth is calculated to create an error map. This map serves as the target for the OCENet. By learning to predict its own errors, the OCENet identifies regions of high aleatoric uncertainty, such as complex boundaries or thin structures. This confidence information is then fed back into the training process to weight the loss function, forcing the detector to focus more on difficult, low-confidence regions.
3. Experiments and Results
The model was evaluated on four major COD benchmarks: COD10K, CAMO, CHAMELEON, and NC4K.
3.1 Quantitative Performance
CANet achieved state-of-the-art results across all datasets. Notably, it demonstrated a significant reduction in Mean Absolute Error (MAE), improving by 14.7% to 18.5% over the previous leading model, SINet. Despite the addition of the uncertainty estimation network, the system remains efficient, with an inference time of approximately 0.03 seconds per image (0.0211s for detection and 0.0094s for confidence estimation).
3.2 Qualitative Analysis
Visual results (Figure 1) demonstrate that the confidence map effectively highlights object boundaries and ambiguous regions. By incorporating this uncertainty, the model produces sharper boundaries and is more successful at identifying thin appendages (e.g., lizard legs) that are often missed by standard models. Furthermore, the confidence network can evaluate prediction accuracy in real-time without requiring ground truth, which is a significant advantage for practical deployment.
4. Discussion and Conclusion
This work highlights that for camouflaged object detection, modeling aleatoric uncertainty is more critical than modeling epistemic uncertainty (model ignorance). By explicitly accounting for label noise and visual ambiguity, the CANet framework achieves superior robustness and boundary accuracy.
Future research directions include exploring the combination of aleatoric and epistemic uncertainty for even more reliable systems, as well as testing the framework with more advanced backbones like Vision Transformers (ViTs). The current success of CANet provides a strong foundation for more transparent and accurate camouflaged object detection in real-world scenarios.
Summary updated