New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Performance degradation #8
Comments
Hi, |
Hi bonlime, Thanks for the quick response. Actually I used your conversion to convert Tensorflow deeplabv3 trained on CityScapes, and measure the mIoU on CityScapes validation sets. They reported ~80% mIoU but I only got ~67%. Is there something I need to be aware of if I use your conversion on CityScapes pretrained model? I will compare the output of the same image, brilliant strategy. Thanks |
How do you preprocess your image? The right way to do it is as follows:
|
I used the same as you did, but cityscapes they use original cityscapes image size (1025, 2049) in evaluation, so what I did is only to do /127.5 - 1. # From tensorflow/models/research/ No worries, let me quickly check the output mask and get back to you. Thanks |
--decoder_output_stride=4 |
I didn't. Could you elaborate a little more why?
|
decoder_output_stride = 4 means they use only one x4 (or x2 in case of OS=8) bilinear upsample, but don't use the last one. |
Hi bonlime, I notice that deeplab_demo.ipynb doesn't do input preprocessing. Am I correct?
Or they did it in tensorflow graph? |
I run both model with my metrics and got the same output, seems my mIoU has some issue and doesn't match the one used by Tensorflow team. |
The issue I found out is that: |
Nice! So without unknown area, mIoU is the same? |
Yes, the same or just 1~2% lower than they reported depends on if implement the input preprocessing exactly as they do. |
Hi,
I'm using the converted pretrain-weights to measure mIoU and observed 10% drop. Could you sure you measurement results, did you still get 84% mIoU on PASCAL using the transferred model and weights?
Thanks
The text was updated successfully, but these errors were encountered: