SDEA block addapted from "Suppressing Features That Contain Disparity Edge For Stereo Matching"
X. Ai, Z. Yang, W. Yang, Y. Zhao, Z. Yu and F. Li, "Suppressing Features That Contain Disparity Edge For Stereo Matching," 2020 25th International Conference on Pattern Recognition (ICPR), 2021, pp. 7985-7991, doi: 10.1109/ICPR48806.2021.9413024.
Implemented according to information from the paper, but according to the parameter numbers for PSMNet implementation, it's still missing something.
Network | Block count | Parameters | Parameters in paper | Missing |
---|---|---|---|---|
SDEA-1 | 3 | 5225027 | 5225158 | 131 |
SDEA-Net | 6 | 5225414 | 5225548 | 134 |
SDEA-2 | 9 | 5225513 | 5225650 | 137 |
Replacing layer G2 and scaling channel with Conv2D_BatchNorm instead of singular Conv2D fixes it, but it's not in the document.