image-segmentation-keras
image-segmentation-keras copied to clipboard
divamgupta
More Pretrained Models
Hi @divamgupta ;
Thank you for your great contribution! Everything works fine for me. This repo helps me a lot.
However, I noticed that there are only pretrained models for pspnet. I am wondering are there any more pretrained models(i.e. fcn pretrained model, segnet pretrained on ADE20K). Cause I am working on some project needed to compare performance. While train all models by my own cost too much time. I will be very grateful if you can release more pretrained models!
Meanwhile, I use a image from ADE20K
and use the pretrained model as described in readme.
here is my code:
`from keras_segmentation.pretrained import pspnet_50_ADE_20K
model=pspnet_50_ADE_20K()
out = model.predict_segmentation(
inp="./input_image.jpg",
out_fname="out.png"
)`
get th predictions as follows:
is that the correct results?
I also use the model.evaluate_segmentation on ADE_20K valdataset.
but get the following results:
'frequency_weighted_IU': 0.008458777348417275, 'mean_IU': 0.0008622930610855478,
I didn't change other things I think. Why this happened?
The low IoU score could be because the class index of the pre-trained model could be different than the class numbers of the pre-processed ADE dataset.
The low IoU score could be because the class index of the pre-trained model could be different than the class numbers of the pre-processed ADE dataset.
Thank you for your reply. I think that is the reason. However, how to match two different class numbers?(How to get the class numbers of pretrained model? )
@divamgupta and more questions, I noticed that in the original UNET. The Encoder process is : conv1-1 conv1-2
conv2-1 conv2-2 ...... but in your implementation(vanilla encoder), the process is conv1-1
conv2-1 ...... seems in each layer, a convolutional block is missing. And this also happens in decoder process. I am wondering why is that?
@divamgupta @fracpete @bakwc @thomasaarholt Anybody could help me on that?😂