README.md

Toward Realistic Camouflaged Object Detection: Benchmarks and Method

Datasets with BBox and Classes Link (First Version)

Google

Baidu Extract Code: 93yd

The first version of the datasets contain category and bounding box annotations.

Datasets	Categories	Training Images	Test Images
COD10K-D	68	6000	4000
NC4K-D	37	2863	1227
CAMO-D	43	744	497

Benchmarks of GLIP and Groudning DINO on COD10K-D, NC4K-D, and CAMO-D datasets (First Version)

Caption: Performance of various detection methods on COD10K-D, NC4K-D, and CAMO-D datasets. The results for a single seed are presented here. The paper presents 10 random seed results.

Generic Methods

Method	Backbone	mAP	AP50	AP75	APm	APl	mAP	AP50	AP75	APm	APl	mAP	AP50	AP75	APm	APl
		COD10K-D					NC4K-D					CAMO-D
YOLOv7-L	CSPDarknet	3.8	8.2	2.8	1.0	4.0	6.8	14.2	6.0	1.7	7.3	5.4	10.2	5.5	8.5	5.4
Faster R-CNN	ResNet-50	8.3	21.0	5.0	4.8	8.8	19.2	39.7	16.0	8.1	20.2	4.7	12.4	2.2	3.5	5.3
YOLOv8-L	CSPVoVNet	9.7	16.8	9.4	2.6	10.4	23.5	34.9	25.2	10.6	24.7	25.4	37.2	26.1	14.4	26.7
Faster R-CNN	ResNet-101	10.8	24.4	7.7	9.2	11.6	23.0	47.2	20.1	10.4	24.0	9.3	21.1	6.9	9.9	10.1
Def-DETR	ResNet-50	12.2	23.1	11.4	6.5	13.1	27.4	49.6	27.9	14.0	29.7	13.3	26.9	12.4	9.7	14.1
Def-DETR	ResNet-101	13.5	23.7	13.5	9.2	14.6	30.9	54.4	32.0	12.4	32.5	13.7	27.4	12.7	13.2	15.5
Cascade R-CNN	ResNet-101	15.3	27.4	15.9	8.5	16.4	27.5	46.9	28.9	11.3	29.2	14.0	26.6	13.1	13.1	14.8
Faster R-CNN	Swin-T	16.3	35.3	13.1	8.6	17.4	29.1	58.8	25.6	16.8	30.4	11.3	32.3	5.5	8.6	12.0
Faster R-CNN	Swin-L	32.1	54.6	33.1	17.1	33.9	49.1	75.8	55.1	22.7	51.3	34.2	67.4	30.2	24.0	36.1

Large Vision-Language Models

Method	Backbone	mAP	AP50	AP75	APm	APl	mAP	AP50	AP75	APm	APl	mAP	AP50	AP75	APm	APl
		COD10K-D					NC4K-D					CAMO-D
GLIP	Swin-T	26.4	36.3	28.5	14.7	28.0	49.6	63.7	53.4	23.9	51.6	32.6	42.9	33.6	40.9	35.1
GLIP + CAFR	Swin-T	28.8	38.2	31.0	16.4	30.6	53.3	67.7	55.2	27.0	55.7	34.7	43.5	34.3	36.9	38.5
GLIP	Swin-L	40.2	47.9	43.5	24.7	42.3	76.9	86.9	80.9	50.4	78.5	63.0	74.4	68.1	52.4	66.8
GLIP + CAFR	Swin-L	42.9	50.0	45.8	28.3	43.0	78.7	89.9	83.9	53.5	80.5	63.6	77.3	70.5	50.0	69.8
GDino	Swin-T	44.8	56.0	47.9	23.5	47.8	69.8	81.0	72.1	37.5	72.4	48.0	59.1	52.4	40.7	52.2
GDino + CAFR	Swin-T	48.5	60.7	51.9	28.7	49.7	72.3	82.7	74.5	35.6	74.7	50.7	60.7	55.3	45.0	55.2
GDino	Swin-B	58.7	70.9	63.1	23.6	62.3	79.9	90.5	84.6	54.8	81.5	68.6	80.6	75.1	55.7	73.0
GDino + CAFR	Swin-B	62.3	74.8	68.3	35.4	67.1	81.9	92.9	87.3	58.1	83.7	72.1	83.3	77.2	56.9	74.5

Datasets with BBox, Classes and Languages Link (Second Version)

New datasets will be provided after the paper is published.

Legend:

• B-Boxes: Bounding Boxes
• Pos-Classes: positive classes
• Tr-Samples: training samples
• Te-Samples: Test Samples

Dataset Type	Datasets	Category	Box	Description	B-Boxes	Pos-Classes	Languages	Tr-Samples	Te-Samples
COD	CHAMELEON				-	-	-	-	76
COD	CAMO	✓			-	8	-	1,000	250
COD	NC4K				-	-	-	-	4,121
COD	COD10K	✓	✓		5,899	69	-	6,000	4,000
RCOD	COD10K-D	✓	✓	✓	11,684	81	10,798	6,172	5,734
RCOD	RCOD-D	✓	✓	✓	12,955	59	11,850	4,192	5,846

Framework install

Our code is based on MMDetection. Here, for the convenience of readers, we have uploaded the full code of mmdetection and our code. If the relevant environment for mmdetection is configured on your server, you can download and use it directly. MMDetection is an open source object detection toolbox based on PyTorch. We adopt MMDetection as our baseline framework from MMdetection

Our environmental installation

• Linux with Python >= 3.10
• conda create -n RCOD python==3.10
• conda activate RCOD
• PyTorch >= 2.1.1 & torchvision that matches the PyTorch version.
• Our CUDA is 11.8
• Install PyTorch 2.1.1 with CUDA 11.8

  conda install pytorch==2.1.1 torchvision==0.16.1 torchaudio==2.1.1 pytorch-cuda=11.8 -c pytorch -c nvidia
  

• pip install mmcv>=2.2.0
• pip install -r requirements/build.txt
• pip install -v -e .

Training on CAFR

• We provide the config files of the three datasets together, thus the number of categories in the config file and the path of the dataset needed to be changed during training. Here, data modification includes:

  RCOD/mmdet/datasets/coco.py  
  RCOD/configs/_base_/coco_detection.py

• We use GLIP+APG as an example to show the training processing:

  CUDA_VISIBLE_DEVICES=0,1,2,3 bash ./tools/dist_train.sh "--config configs/glip/glip_swin_tiny_cafr.py --work-dir /home/output 4
  

Citation

If you use this toolbox or benchmark datasets in your research, please cite this project.

@article{rcod,
	title={Toward Realistic Camouflaged Object Detection: Benchmarks and Method},
	author={Xin, Zhimeng and Wu, Tianxu and Chen, Shiming and Ye, Shuo and Xie, Zijing and Zou, Yixiong and You, Xinge and Guo, Yufei},
	journal={arXiv preprint arXiv:2501.07297},
	year={2025}
}