给定一张输入图像,找出图中的伪装色目标,并输出视觉显著注意力图。
本模型是针对伪装色目标(颜色、纹理等统计特征与所处环境一致)进行定位,并输出视觉区域图,模型是在学术数据集上训练,需要待检测图像目标类别在数据集范围之内。
在ModelScope框架上,提供输入图片,即可以通过简单的Pipeline调用使用当前模型。
from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks
from modelscope.outputs import OutputKeys
camouflag_detect = pipeline(Tasks.semantic_segmentation, model='damo/cv_res2net_camouflaged-detection')
img_path ='https://modelscope.oss-cn-beijing.aliyuncs.com/test/images/image_camouflag_detection.jpg'
result = camouflag_detect(img_path)
import cv2
cv2.imwrite('./result.jpg',result[OutputKeys.MASKS])
DataSet | MAE | Sm | maxF~β | F^w~β |
---|---|---|---|---|
CAMO | 0.068 | 0.838 | 0.823 | 0.780 |
CHAMELEON | 0.026 | 0.907 | 0.888 | 0.858 |
COD10K | 0.031 | 0.843 | 0.789 | 0.739 |
NC4K | 0.042 | 0.862 | 0.841 | 0.797 |
@article{gao2019res2net,
title={Res2Net: A New Multi-scale Backbone Architecture},
author={Gao, Shang-Hua and Cheng, Ming-Ming and Zhao, Kai and Zhang, Xin-Yu and Yang, Ming-Hsuan and Torr, Philip},
journal={IEEE TPAMI},
year={2021},
doi={10.1109/TPAMI.2019.2938758},
}