niecongchong/hrnet-keras-semantic-segmentation

keras-HRNet

hrnet-keras-semantic-segmentation's Issues

hrnet performance issue

hello author,

I found that the HRNET performance is 10% lesser than UNET or FCN model. I have used 96x96 image. Is this because i choose low resolution image for training?

Thank You

weight of keras version

Hello, thanks for your code! Could you please share the weight of keras version ( model.load_weights('seg_hrnet-08-4.2117-0.9428-0.4832.hdf5')) with me? Thank you very much

Choice of kernel initializer

Any reason of why you're using kernel_initializer='he_normal' in the convolution layers ?

License information

Hi,

Thanks for this example code. Can you specify the license with which you are presenting this? MIT/Apache?

Thank you.

HRNetV2-W18-SMALL-v2

I wish to use HRNetV2-W18-SMALL-v2 for semantic segmentation, is it possible for you to provide that implementation?

I tried the official HRNET library, but am facing installation issues

Thanks!

Is the output shape same as the original pytorch HRNet which is (1,19,original_height/4, original_weight/4) ?

Hello, I am working on HRNet these days, and I found that in your test code, you use two images as one input to run the model, and it seems that you did not do the upsample process compared to original pytorch model. Because in the original model, for maintaining high resolution, the model's output shape is one fourth of the original image. And we have to do upsample process to make it be the original size.
But in your code, I did not see this step, so I wonder how you build the model ? Are there some differences ? Besides, why do you put two images together into the model ?

pretrained network weights

Hello, can you share the weights of your pretrained network?

Repetition of the blocks in stages 3 and 4

According to the paper, the blocks in stages 3 and 4 should be repeated 4 and 3 times respectively. I can do a PR to fix it if you want.