lsa1997
commented
2 years ago Hi, I have updated these networks and you can check them here.
Open AlphaNext opened 2 years ago
AlphaNext
commented
2 years ago 以下是我针对Tabel3 的复现结果，在+PCAA时，与report的出入比较大，并不是最好的，请作者指点一二。
注： 以下结果都是使用仓库源代码，且训练的机器环境完全一致，结果都取的是日志文件中的best_IU一项
| Method | Backbone | mIoU |
|---|---|---|
| FCN(Baseline) | ResNet-50 | 73.50 |
| +ASPP(deeplap-v3) | ResNet-50 | 78.27 |
| +NL(nlnet) | ResNet-50 | 79.07 |
| +OCR(ocrnet) | ResNet-50 | 77.93 |
| +PCAA(caanet) | ResNet-50 | 78.50 |
| ResNet-101 backbone的结果： | Method | Backbone | mIoU |
|---|---|---|---|
| FCN(Baseline) | ResNet-101 | 75.31 | |
| +ASPP(deeplap-v3) | ResNet-101 | 79.99 | |
| +NL(nlnet) | ResNet-101 | 79.98 | |
| +OCR(ocrnet) | ResNet-101 | 79.81 | |
| +PCAA(caanet) | ResNet-101 | 78.86 |
lsa1997
commented
2 years ago Can you provide your training settings and PyTorch environments? Empirically, performance on Cityscapes val is not very stable. However, the results you provided seem much lower than the average. FYI, the average result of my recent 5 models (caanet, res50) is 79.29 +- 0.14 mIoU.
感谢作者的工作，有个小问题，Table3 在Citycapes的实验结果，使用哪个代码训练的FCN（baseline）、FCN+NL、FCN+OCR、FCN+PCAA等