Our model ranks first in the challenging SOC benchmark up to now (2019.11.6).
python2.7, pytorch 0.4.0
Modify the pathes of backbone and datasets, then run test_CPD.py
VGG16 backbone: google drive, BaiduYun (code: gb5u)
ResNet50 backbone: google drive, BaiduYun (code: klfd)
VGG16 backbone: google drive
ResNet50 backbone: google drive
Maximum F-measure
| Model | FPS | ECSSD | HKU-IS | DUT-OMRON | DUTS-TEST | PASCAL-S |
|---|---|---|---|---|---|---|
| PiCANet | 7 | 0.931 | 0.921 | 0.794 | 0.851 | 0.862 |
| CPD | 66 | 0.936 | 0.924 | 0.794 | 0.864 | 0.866 |
| PiCANet-R | 5 | 0.935 | 0.919 | 0.803 | 0.860 | 0.863 |
| CPD-R | 62 | 0.939 | 0.925 | 0.797 | 0.865 | 0.864 |
MAE
| Model | ECSSD | HKU-IS | DUT-OMRON | DUTS-TEST | PASCAL-S |
|---|---|---|---|---|---|
| PiCANet | 0.046 | 0.042 | 0.068 | 0.054 | 0.076 |
| CPD | 0.040 | 0.033 | 0.057 | 0.043 | 0.074 |
| PiCANet-R | 0.046 | 0.043 | 0.065 | 0.051 | 0.075 |
| CPD-R | 0.037 | 0.034 | 0.056 | 0.043 | 0.072 |
pre-computed maps: google drive
BER
| Model | SBU | ISTD | UCF |
|---|---|---|---|
| DSC | 5.59 | 8.24 | 8.10 |
| CPD | 4.19 | 6.76 | 7.21 |
@InProceedings{Wu_2019_CVPR,
author = {Wu, Zhe and Su, Li and Huang, Qingming},
title = {Cascaded Partial Decoder for Fast and Accurate Salient Object Detection},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2019}
}