DQnet: Cross-Model Detail Querying for Camouflaged Object Detection

This is the official implementaion of paper DQnet: Cross-Model Detail Querying for Camouflaged Object Detection.

Illustration

With the contextual representation of ViT as global cues, our model queries crucial local details from the multi-scale CNN features. The enhanced representations have clear boundaries as well as few background noise, corresponding well with underlying camouflaged objects.

Updates

• 2022-11-11
  • Finished our paper.

Model

The model used pretrained weights for ResNet50 and ViT. They can be found at:

• Resnet50
• ViT

Usage

First clone the repository locally:

git clone https://github.com/CVPR23/DQnet.git

Second download the weights for ResNet50 and ViT.

Some core dependencies:

• timm == 0.4.12
• torch == 1.11.0

More details can be found in <./requirements.txt>

Datasets

More details can be found at:

• COD Datasets
• CAMO Datasets
• NC4K Datasets
• CHAMELEON Datasets

For training:

You can use our default configuration, like this:

$ python main.py --model-name=DQnet --config=configs/DQnet/DQnet.py --datasets-info ./configs/_base_/dataset/dataset_configs.json --info demo

You can also use :

$ sh train.sh

For testing:

You can use our default configuration, like this:

$ python test.py 

You can also use :

$ sh test.sh

Paper Details

Method Detials

DQNet

RBQ

Comparison

Visualization of mutil-scale details querying