search

clovaai / ext_portrait_segmentation

MIT License
239 stars 68 forks source link

Accuracy issue, implementation difference compared to paper. #17

Open avinash31d opened 4 years ago

avinash31d commented 4 years ago

Hi,

I tried retrain the model on biadu_V2, i am getting training iou as 0.945, but i evaluated on real life images, I am not able to get good results.

Also I can see the loss function you used in here is lovasz whereas the paper SINET has a completely different loss function. Can you provide some information on why the loss function and what could be the problem with these test images?

HYOJINPARK commented 4 years ago

@avinash31d In here, I used normal CE loss and you can change it by modifying json file

avinash31d commented 4 years ago

@HYOJINPARK

Thanks for taking out time,

I tried CE but it seems lovasz does it better. Also I posted few results just to compare. Can you crop the left side image and run through your model and upload the results here for comparison because I doubt if my model is working.

HYOJINPARK commented 4 years ago

@avinash31d
You can use my model, I uploaded the pth file

ZshahRA commented 4 years ago

Hello @avinash31d,

Could you please provide your code for visualize_video.py. I want to see how you test your new images. I have trouble in loading Extremec3Net model in visualize_video.py. Thank you