Road segmentation with PyTorch

My work with Pytorch -
Trained on custom dataset and model.
Github link - project is still not finished !


Amazing work!! I have a few questions:

  1. Does it work in real-time?
  2. Can this model generalize well to sky images? I am guessing it will but wanna know what are your thoughts on it.
  3. How long did it took to train and which GPU use used?
1 Like

Hello thank you !

Yes yesterday I converted this model to Onnx and it works pretty fast Real Time.

Yes this model will work very good on sky images. - You will need a lot of training data, but I don’t think it’s as much as I do, because the sky doesn’t have as many forms as roads. I think 500 images, cloudy/clear sky/mix will be enough.

I am using RTX 3060 12GB and it took at least 12 hours for good results.


(post deleted by author)

Thanks man! Btw is Tesla P100 fast or RTX 3060? I couldn’t find the TFlops specs to compare them on NVIDIA website.

I am not sure, sorry.

Okay, it’s fine. Thanks for code anyways!

tesla p100 has 9.3 TFLOPS on single precision vs 3060 has 12.8 on single precision


Great! Thanks for the information @arya47.