Camouflage-Aware Low-Light Enhancement and Cross-Attentional Mamba Fusion for RGB-P Camouflaged Object Detection (Displays 2026)
May 28, 2026 · View on GitHub
Authors: Weiyun Liang, Chunyuan Chen, Jing Xu, Bin Wang, and Donglin Wang
1. Overview
- This repository provides code for "Camouflage-Aware Low-Light Enhancement and Cross-Attentional Mamba Fusion for RGB-P Camouflaged Object Detection", Displays, 2026. Paper
1.1. Introduction
RGB-P camouflaged object detection (COD) aims to identify objects that seamlessly blend into their surroundings by leveraging complementary cues from paired RGB and polarization images. However, existing RGB-P COD methods suffer from significant performance degradation under low-light conditions due to the low quality of polarization maps, which contain limited camouflaged object cues. Moreover, while effectively mining complementary discriminative information from multi-modal features is particularly prominent in low-light scenes, existing RGB and polarization fusion strategies often focus only on coarse feature-level fusion, lacking fine-grained alignment between the two feature spaces. To address these issues, we propose a Camouflage-aware low-light enhancement and Cross-attentional Mamba fusion network, namely C2Net, for RGB-P COD. Specifically, a camouflage-aware low-light enhancement (CaLLE) module is introduced to inject camouflage-aware semantics into the low-light enhancement process, highlighting camouflaged objects under low-light conditions. In addition, a cross-attentional Mamba fusion (CAMF) module is proposed to leverage Mamba’s state space modeling capability to perform fine-grained alignment between multi-modal feature state spaces, enabling more effective RGB-P information fusion. Extensive experiments demonstrate the proposed method achieves superior performance compared to state-of-the-art methods on four RGB-P and RGB COD benchmark datasets.
1.2. Framework
Figure 1: The overall architecture of our proposed C2Net, which comprises two components: a camouflage-aware low-light enhancement (CaLLE) module and a Detector. The CaLLE module utilizes camouflage-aware reverse attention module (CRAM) to inject camouflage-aware semantics into an LLE process. The Detector utilizes a two-stream encoder to generate enhanced RGB and polarization features, four cross-attentional Mamba fusion (CAMF) modules to fuse multi-modal features, and three ConvBlocks to aggregate multi-stage features in a top-down manner.
1.3. Quantitative Results
Figure 2: Quantitative results of our method and the SOTA methods on RGB-P and RGB benchmark datasets.
1.4. Qualitative Results
Figure 3: Qualitative Comparison.
2. Proposed Baseline
2.1. Prepare the Data
The training and testing datasets can be downloaded from https://github.com/cvhfut/PCOD_1200 and https://github.com/GewelsJI/SINet-V2/tree/main .
You can modify config.py to set all the data paths.
2.2 Training Configuration
-
Traning hyperparameters and data paths can be modified in
config.py. -
Pretrained weights for PVTv2 backbone can be downloaded from https://github.com/whai362/PVT/tree/v2/classification .
-
Installing necessary packages:
- pytorch: https://pytorch.org
- pysodmetrics: https://github.com/lartpang/PySODMetrics
- VMamba: https://github.com/zifuwan/Sigma
2.3 Training, Inference, and Testing
-
Run
train.pyto train the model. -
Run
inference.pyto generate the prediction maps. Set path to your trained checkpoint (pth_path). -
Run
evaluate.pyto calculate the evaluation metrics for the prediction maps. Set path to your predictions (pred_path).
3. Pre-computed Maps
Google Drive or Baidu Drive (extraction code: hraa)
4. Citation
Please cite our paper if you find the work useful, thanks!
@article{LIANG2026103265,
title = {Camouflage-aware low-light enhancement and cross-attentional Mamba fusion for RGB-P camouflaged object detection},
journal = {Displays},
volume = {91},
pages = {103265},
year = {2026},
author = {Weiyun Liang and Chunyuan Chen and Jing Xu and Bin Wang and Donglin Wang},
}
Acknowledgement
The code is built on Sigma, Zero-DCE, and SINetV2. Please also follow the corresponding licenses. Thanks for their awesome work.