search

HRNet / HRNet-Semantic-Segmentation

The OCR approach is rephrased as Segmentation Transformer: https://arxiv.org/abs/1909.11065. This is an official implementation of semantic segmentation for HRNet. https://arxiv.org/abs/1908.07919
Other
3.17k stars 690 forks source link

Problem in mIoU #228

Open bravozyz opened 3 years ago

bravozyz commented 3 years ago

Hi~ I'm trying to transform HRNet for semantic segmentation into other deep learning framework, MindSpore, but I got problems in mIoU result.

Dataset and Configurations: I use Cityscapes as training and evaluation dataset. All settings are same as yours.

Problems: image This is my result. IoU results of some classes are too low, so that the average, 0.68, can't achieve the target.

I have tried a lot, such as changing interpolation method in data augmentation from bilinear in nearest, increasing the scaling factor, modifing weights in loss function, changing learning rate. However none of these worked, neither increasing nor decreasing my result.

Are there any other factors contributing to the present results? Can you give me some other suggestions? Thank you very much!

PkuRainBow commented 3 years ago

@bravozyz According to your results, I guess the main reason might be failing to load the ImageNet pre-trained weights correctly.