Code repository for our paper entilted "Calibrated RGB-D Salient Object Detection" accepted at CVPR 2021.
- pytorch 1.0.0+; torchvision; PIL; numpy
Our saliency maps.
【1】Saliency Maps, (fetch code is j93d), by our DCF trained on NJUD & NLPR (2185).
【2】Saliency Maps, (fetch code is aeq0), by our DCF trained on NJUD & NLPR & DUT (2985).
- Notice that, the depth map is unified, which means that the closer region is closer to 1, and the farther region is closer to 0. The testset results of the new dataset ReDWeb-S can be downloaded in here (fetch code is likm).
Our pre-trained model for inferring your own dataset.
【1】Download the pre-trained model, (fetch code is 2t7g), which is trained on NJUD & NLPR & DUT. Or the another model, (fetch code is epp9), which is trained on NJUD & NLPR.
【2】Set the data path and ckpt_name in demo_test.py, correctly.
【3】Run python demo_test.py to obtain the saliency maps.
【1】Stage 1: Run python demo_train_pre.py, which performs the Depth Calibration Strategy.
【2】Stage 2: Run python demo_train.py, which performs the Fusion Strategy.
-
The related all test datasets in this paper can be found in this link (fetch code is b2p2).
-
This evaluation tool is used to evaluate the above saliency maps in this paper.
-
The training set used in this paper can be accessed in (NJUD+NLPR), code is 76gu and (NJUD+NLPR+DUT), code is 201p.
We thank all reviewers for their valuable suggestions. At the same time, thanks to the large number of researchers contributing to the development of open source in this field, particularly, Deng-ping Fan, Runmin Cong, Tao Zhou, etc.
Our feature extraction network is based on CPD backbone.
@InProceedings{Ji_2021_DCF,
author = {Ji, Wei and Li, Jingjing and Yu, Shuang and Zhang, Miao and Piao, Yongri and Yao, Shunyu and Bi, Qi and Ma, Kai and Zheng, Yefeng and Lu, Huchuan and Cheng, Li},
title = {Calibrated RGB-D Salient Object Detection},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2021},
pages = {9471-9481}
}
If you have any questions, please contact us ( wji3@ualberta.ca ).