The result of HRNet+OCR

[ydhongHIT]

I can't find the val result of 84.9%mIoU in the OCR paper. Is this the result reproduced by yours? If so ,can you describe your implementation details?

[khoshsirat]

This result is reproduced by the following paper:
Hierarchical Multi-Scale Attention for Semantic Segmentation

And you can download their pretrained weights from here:
https://github.com/NVIDIA/semantic-segmentation