Best available semantic segmentation in PyTorch


(Pete Florence) #1

Hey folks –

I’m looking for the best semantic segmentation network I can find that is available in PyTorch. (Best as measured by mean IoU on Cityscapes / PASCAL VOC2012)

The best number I can find in an available repo is in this implementation from the authors of Dilated Residual Networks, which in their readme they say can achieve 76.3% IoU on Cityscapes. Also noteworthy are some implementations of DeepLab ResNet which both claim mid-70s on PASCAL VOC, here is one. (Cityscapes numbers are typically lower though, so the DRN performance on Cityscapes is significantly higher.)

Although I’ve come across many attempted reimplementations of semantic segmentation networks in PyTorch, many are either missing the original paper’s numbers by a large margin or are otherwise not reporting IoU numbers.

Has anybody else found any other noteworthy semantic segmentation implementations available in PyTorch I should be looking at?

Thanks!
Pete


(Manu Mathew) #2

I found the following interesting:


Various semantic segmentation networks with mIoU ranging up to 80.13 on Cityscapes are listed.


#4

I’m also looking for such implementations. Are there any recent implementations for the new PyTorch version?


(Sanyam Rajpal) #5

The list might be too generic, but it probably has more than one could ask for in semantic segmentation.