search

qubvel / segmentation_models

Segmentation models with pretrained backbones. Keras and TensorFlow Keras.
MIT License
4.73k stars 1.03k forks source link

Pixelated output from PSPNET #435

Open Hamid-Sarraf opened 3 years ago

Hamid-Sarraf commented 3 years ago

I'm trying to apply PSPNet to segment a set of Transmision Xray Images (similar to medical x-ray images but for materials science). I've trained the network with a set of Ground-Truth images but getting the very pixelated segmented images as output (see the attached). Although the mIoU is high during training, still the test results are pixelated. I changed the default down-sampling from 8 to 4 and it became a bit better but still pixelated.

I'll appreciate it if you please help me with what other parameters for PSPNet I can change to get high res results like another network (Unet, Unet++, Nestnet) I used from this repository. Thanks, PSPNet

khanhak commented 2 years ago

@Hamid-Sarraf Did you ever resolve this issue? I'm having the same problem...

Hamid-Sarraf commented 2 years ago

Unfortunately No.

khanhak commented 2 years ago

For anyone else reading this, I was able to get significantly better per-pixel accuracy by using Unet from this library instead of pspnet.

On Mar 19, 2022, at 10:40 AM, Hamid Torbati-Sarraf @.***> wrote:

 Unfortunately No.

— Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you commented.