Segment, Magnify and Reiterate: Detecting Camouflaged Objects the Hard Way. Jia Qi and Yao Shuilian and Liu Yu and Fan Xin and Liu Risheng and Luo Zhongxuan. CVPR2022.

paper download

The training and testing experiments are conducted using PyTorch with a single Tesla V100 GPU of 36 GB Memory.

Note that SegMaR is only tested on Ubuntu OS with the following environments.

• Creating a virtual environment in terminal: conda create -n SegMaR python=3.6.

• Installing necessary packages: pip install -r requirements.txt.

• Installing NVIDIA-Apex (Under CUDA-10.0 and Cudnn-7.4).

• Installing MobulaOP for Sampler operation.

  # Clone the project
  git clone https://github.com/wkcn/MobulaOP
  
  # Enter the directory
  cd MobulaOP
  
  # Install MobulaOP
  pip install -v -e .
  
  

• Downloading training dataset (COD10K-train) and move it into ./OurModule/datasets/train/.

• Downloading testing dataset (COD10K-test + CAMO-test + CHAMELEON) and move it into ./OurModule/datasets/test/.

• You can use discriminative mask google drive link or discriminative mask baidu drive link, code e5ym, or run ./OurSampler/DiscriminativeMask.py to generate your discriminative mask. If you use a dataset without this mask (like COCO or Pascal VOC), you can replace discriminative mask with binary groundtruth.

• After you download all the training datasets, just run ./OurModule/train.py to generate the model.

• For iterative training: generator.load_state_dict(torch.load('./OurModule/models/xxx.pth')).

• For the first stage pretrained model, code y4v3, or google drive link

• After you download all the pre-trained model and testing datasets, just run ./OurModule/test.py to generate the prediction map. Your save directory is ./OurModule/results.py.

• Test results, code pxu7

• New NC4K results or baidu drive link, code naif

• Make sure that you have installed MobulaOP in your virtual environment.

• For sampler operation, just run ./OurSampler/Sampler_Distort.py.

• For restoration operation, just run ./OurSampler/Sampler_Restort.py.

• For the directory of original prediction or restoration prediction, please see our codes details.

• One-key evaluation is written in MATLAB code, please follow this the instructions in main.m and just run it to generate the evaluation results.

@InProceedings{Jia_2022_CVPR,
    author    = {Jia, Qi and Yao, Shuilian and Liu, Yu and Fan, Xin and Liu, Risheng and Luo, Zhongxuan},
    title     = {Segment, Magnify and Reiterate: Detecting Camouflaged Objects the Hard Way},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {4713-4722}
}