(ECCV 2024) Open-Vocabulary Camouflaged Object Segmentation

@inproceedings{OVCOS_ECCV2024,
  title={Open-Vocabulary Camouflaged Object Segmentation},
  author={Pang, Youwei and Zhao, Xiaoqi and Zuo, Jiaming and Zhang, Lihe and Lu, Huchuan},
  booktitle=ECCV,
  year={2024},
}

Note

Details of the proposed OVCamo dataset can be found in the document for our dataset.

Prepare Dataset

1. Prepare the training and testing splits: See the document in our dataset for details.
2. Set the training and testing splits in the yaml file env/splitted_ovcamo.yaml:
   • OVCamo_TR_IMAGE_DIR: Image directory of the training set.
   • OVCamo_TR_MASK_DIR: Mask directory of the training set.
   • OVCamo_TR_DEPTH_DIR: Depth map directory of the training set. Depth maps of the training set which are generated by us, can be downloaded from https://github.com/lartpang/OVCamo/releases/download/dataset-v1.0/depth-train-ovcoser.zip
   • OVCamo_TE_IMAGE_DIR: Image directory of the testing set.
   • OVCamo_TE_MASK_DIR: Mask directory of the testing set.
   • OVCamo_CLASS_JSON_PATH: Path of the json file class_info.json storing class information of the proposed OVCamo.
   • OVCamo_SAMPLE_JSON_PATH: Path of the json file sample_info.json storing sample information of the proposed OVCamo.

Training/Inference

1. Install dependencies: pip install -r requirements.txt.
   1. The versions of torch and torchvision are listed in the comment of requirements.txt.
2. Run the script to:
   1. train the model: python .\main.py --config .\configs\ovcoser.py --model-name OVCoser;
   2. inference the model: python .\main.py --config .\configs\ovcoser.py --model-name OVCoser --evaluate --load-from <path of the local .pth file.>.

Evaluate the Pretrained Model

1. Download the pretrained model.
2. Run the script: python .\main.py --config .\configs\ovcoser.py --model-name OVCoser --evaluate --load-from model.pth.

Evaluate Our Results

1. Download our results and unzip it into <path>/ovcoser-ovcamo-te.
2. Run the script: python .\evaluate.py --pre <path>/ovcoser-ovcamo-te

LICENSE

• Code: MIT LICENSE
• Dataset:

  OVCamo by Youwei Pang, Xiaoqi Zhao, Jiaming Zuo, Lihe Zhang, Huchuan Lu is licensed under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International