Home

Awesome

Detecting Camouflaged Object in Frequency Domain

An unofficial implementation for Detecting Camouflaged Object in Frequency Domain, CVPR 2022 in PyTorch.

[paper] | [supp].

Requirements

Here, we list our environment for the experiment on both Linux or Window.

# install

python 3.6
torch == 1.3.1
torchvision == 0.4.2
torch-dct == 0.1.5
numpy == 1.19.5
einops == 0.4.1

# for evaluation (optional)

pysodmetrics == 1.3.0

The package torch-dct is used for the differential discrete cosine transformation in PyTorch, and more details can be found in this repo. Note: A higher version for PyTorch has included this function and it may cause some problem. You should modify the source code of torch-dct or our code to solve the problem.

The package pysodmetrics is used for calculating the metrics for camouflaged object detection based on Python, as COD and SOD share similar metrics. The usage of this package can be found in link.

Data

Before testing the network, please download the data:

---- CAMO
   |---- Images
       |---- Test
           |---- ****.jpg
       |---- Train
           |---- ****.jpg
   |---- GT
       |---- ****.png

---- CHAMELEON_TestingDataset
   |---- Image
       |---- ****.jpg
   |---- GT
       |---- ****.png

---- COD10K-v3
   |---- Test
       |---- Image
           |---- ****.jpg
       |---- GT_Object
           |---- ****.png
   |---- Train
       |---- Image
           |---- ****.jpg
       |---- GT_Object
           |---- ****.png

Recommendation: you could extract these data and put them to the same folder (e.g. ./COD_datasets/). Then, the folder should contain three folders: CAMO/, CHAMELEON_TestingDataset/, COD10K-v3/.

Test and Evaluation

Test

It is very simple to test the network. You can follow these steps:

  1. You need to download the model weights [Baidu Yun, qr1n].

  2. Change the output path in train.py Line. 36 to your need. We always use the name of the testing dataset.

36  -   save_supervision_path = os.path.join("results", "COD10K")
    +   save_supervision_path = '/output/path/'
  1. You need to set the path to the testing dataset which you want in data_loader1.py.
53    self.img_dir = 'G:/DataSet/COD10K-v3/Test/Image/'
      self.label_dir = 'G:/DataSet/COD10K-v3/Test/GT_Object/'
      # self.img_dir = '/CAMO/Images/Test/'
      # self.label_dir = '/CAMO/GT/'
      # self.img_dir = '/COD10K-v3/Test/Image/'
      # self.label_dir = '/COD10K-v3/Test/GT_Object/'
      # self.img_dir = '/CHAMELEON_TestingDataset/Image/'
      # self.label_dir = '/CHAMELEON_TestingDataset/GT/'

If you follow the data preparation above, you can simply use the existing code.

  1. Run main.py.

Evaluation

You can use Matlab or Python script for evaluation.

You need to change the path of the ground-truth and the predictions in eval.py. Using the python script is more simple and efficient.

6    # gt_path = 'G:/DataSet/CAMO/GT/'
     # gt_path = 'G:/DataSet/CHAMELEON_TestingDataset/GT/'
     gt_path = 'G:/DataSet/COD10K-v3/Test/GT_Object/'
     predict_path = './results/COD10K/'

Then run python eval.py. You can get the Smeasure, mean Emeasure, weighted Fmeasure, and MAE.

Note: We also upload the results [Baidu Yun, 1mlw].

You can also use the one-key evaluation toolbox for benchmarking provided by Matlab version.

Metrics

Finally, we get the following performance.

<table style="text-align: center"> <tr> <td rowspan='2'></td> <td colspan='4'>COD10K</td> <td colspan='4'>CAMO</td> <td colspan='4'>CHAMELEON</td> </tr> <tr> <td>Sm</td> <td>Em</td> <td>wFm</td> <td>MAE</td> <td>Sm</td> <td>Em</td> <td>wFm</td> <td>MAE</td> <td>Sm</td> <td>Em</td> <td>wFm</td> <td>MAE</td> </tr> <tr> <td>Paper</td> <td>0.837</td> <td>0.918</td> <td>0.731</td> <td>0.030</td> <td>0.844</td> <td>0.898</td> <td>0.778</td> <td>0.062</td> <td>0.898</td> <td>0.949</td> <td>0.837</td> <td>0.027</td> </tr> <tr> <td>Ours</td> <td>0.8404</td> <td>0.9187</td> <td>0.7288</td> <td>0.0297</td> <td>0.8435</td> <td>0.8949</td> <td>0.7746</td> <td>0.0629</td> <td>0.8974</td> <td>0.9497</td> <td>0.8350</td> <td>0.0270</td> </tr> </table>

Citations

If you are using this repo in a publication, please consider citing the origin paper:

@InProceedings{Zhong_2022_CVPR,
    author    = {Zhong, Yijie and Li, Bo and Tang, Lv and Kuang, Senyun and Wu, Shuang and Ding, Shouhong},
    title     = {Detecting Camouflaged Object in Frequency Domain},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {4504-4513}
}

Acknowledgement