Authors: Ranwan Wu, Tian-Zhu Xiang, Guo-Sen Xie, Rongrong Gao, Xiangbo Shu, Fang Zhao, Ling Shao

Welcome to the official PyTorch implementation repository of our paper Uncertainty-Aware Transformer for Referring Camouflaged Object Detection, accepted to IEEE TIP 2025.

Figure.1 Architecture of uncertainty-aware transformer (UAT) for Ref-COD. UAT takes a camouflaged image and several referring images as input, respectively. Next, basic feature extraction on these images is performed. Then, the extracted features are fed into referring feature aggregation (RFA), cross-attention encoder (CAE), and transformer probabilistic decoder (TPD) to integrate visual reference into camouflage feature, aggregate multi-layer camouflage features, and model the dependencies between patches/tokens via Bayesian uncertainty learning, respectively. Finally, the predictions from all four stages are supervised by $L_{total}$ collaboratively.

Python v3.6, Pytorch 0.4.0+, Cuda 10.0, TensorboardX 2.0, opencv-python

• Please visiting RefCOD for training and testing data. Thanks for their contributions.

• Download the training and testing dataset, and place them in the ./dataset floder.
• Download the pre-trained weights of pvtv2[code:2025] on Baidu Netdisk, and place them in the ./pvt_weights floder.
• Run python train.py to train the model.
• You can also download the our pre-trained UAT.pth with access code 2025 on Baidu Netdisk directly.

• After training, run python infer.py to generate the prediction maps of UAT.
• You can also download our prediction maps UAT-Maps[code:2025] on Baidu Netdisk.

• After training, run python test.py to evaluate the performance of UAT.

• Qualitative comparison

Table.1 Quantitative comparison with some SOTA models on referring camouflaged bbject detection benchmark datasets.

This repo is mainly built based on R2CNet. Thanks for the great work! If you have any technical questions, feel free to contact wuranwan2020@sina.com. If our work inspires your research, please cite it and start this project. We appreciate your support!