Weakly-Supervised Camouflaged Object Detection with Scribble Annotations (AAAI23, ORAL)

Authors: Ruozhen He, Qihua Dong, Jiaying Lin, and Rynson Lau (* joint first authors)

Paper Link: arxiv

Dataset

• We relabeled 4,040 images (3,040 from COD10K, 1,000 from CAMO) with scribbles and proposed the S-COD dataset (Download) for training. In our annotations, "1" stands for foregrounds, "2" for backgrounds, and "0" for unlabeled regions. (The image is viewed as black because its range is 0-255)
• Download the training dataset (COD10K-train) at here.
• Download the testing dataset (COD10K-test + CAMO-test + CHAMELEON) at here.

Experimental Results

• We provide CRNet testing maps and training weights presented in the papers.

Code

Requirements

git clone --recurse-submodules https://github.com/dddraxxx/Weakly-Supervised-Camouflaged-Object-Detection-with-Scribble-Annotations.git
pip install -r requirements.txt

Pretrained weights

The pretrained weight can be found here: ResNet-50.

Train

• Download the dataset and pretrained model. (examples of train.txt and test.txt are in the path ./CodDataset)
• Modify the path in train.py.
• Run python train.py.

Test and Evaluate

• The evaluation is done using the submodule PySODEvalToolKit. Add the json files according to its instruction. (examples of json files are in the path ./CodDataset)
• Modify the path and filename.
• Run python test.py.

Credit

The code is partly based on SCWSSOD, GCPANet and GatedCRFLoss.