search

oandrienko / fast-semantic-segmentation

ICNet and PSPNet-50 in Tensorflow for real-time semantic segmentation
220 stars 41 forks source link

Pre-train model for evaluation? #2

Closed Minotaur-CN closed 6 years ago

Minotaur-CN commented 6 years ago

Dear oandrienko,

Thanks for your code ,

  1. Could you upload your trained model for evaluation?
  2. Inference.py:100, output_channels is not defined in code. If setting output_channels=1, the gray result was showed in the result. If setting output_channels!=1. raise TypeError("Cannot handle this data type") on line 102 "im = Image.fromarray(predictions)"

Thanks

oandrienko commented 6 years ago

Hey, thanks for your interest in the project and sorry for the delay.

Thanks for brining up the issue with the inference script. I have updated it so you can now use it with both PSPNet50 or an ICNet model to output either RGB colours or Cityscapes Label IDs.

For the pre-trained ICNet model: I have been away for the last while but am now in the process of figuring out the optimal training and compression setup for ICNet. I can mention you in this issue once I have the ICNet weights uploaded in the near future. In the meantime, I do have the baseline PSPNet50 model up which I have recently uploaded. You can find a link to the trained PSPNet model here. There is also a usage guide for training PSPNet50 here.

Minotaur-CN commented 6 years ago

Thanks for your reply!

oandrienko commented 6 years ago

Hey @Minotaur-CN, I recently uploaded an older compressed ICNet model I trained back in September. If can be found here. A small guide for training can be found here.

The model is ~1.5% mIoU off from the original accuracy reported in the paper. I haven't had much time but am still working on improving the training process. Hope this helps.