给定一张输入图像,通过目标边界信息引导找出图中的显著性目标,并输出视觉显著注意力图。
本模型适用范围较广,预测像素视觉显著注意程度,但不涉及图像中的语义信息。
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
salient_detect = pipeline(Tasks.semantic_segmentation, model='damo/cv_res2net_salient-detection')
img_path ='https://modelscope.oss-cn-beijing.aliyuncs.com/test/images/image_salient_detection.jpg'
result = salient_detect(img_path)
import cv2
cv2.imwrite('./result.jpg',result[OutputKeys.MASKS])
DataSet | MAE | Sm | maxF~β | F^w~β |
---|---|---|---|---|
DUTS-TE | 0.034 | 0.897 | 0.887 | 0.853 |
DUT-OMRON | 0.064 | 0.833 | 0.780 | 0.750 |
SOD | 0.093 | 0.797 | 0.840 | 0.766 |
PASCAL-S | 0.081 | 0.820 | 0.817 | 0.777 |
ECSSD | 0.029 | 0.932 | 0.946 | 0.927 |
HKU-IS | 0.027 | 0.918 | 0.931 | 0.907 |
MSRA10K | 0.056 | 0.893 | 0.899 | 0.871 |
@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},
}