search

alexgkendall / SegNet-Tutorial

Files for a tutorial to train SegNet for road scenes using the CamVid dataset
http://mi.eng.cam.ac.uk/projects/segnet/tutorial.html
847 stars 517 forks source link

High training accuracy, results shifted and noisy #73

Closed MustafaArikan closed 5 years ago

MustafaArikan commented 7 years ago

Dear Alex, dear all,

I am training segnet with multi-channel images using lmdb and the data layer. After a while the training accuracy is quite high (>0.9) and the loss gets low but I am having these strange artifacts in the results.

Top image, GT, Bottom image is one of the results grafik

At first I thought this is a export, shape or data type problem. Then I tried to experiment with the learning rate and let segnet train for more epochs, the results seem to shift and there is less noisy structures but I still don't know what the problem is. Do you have any ideas ?

grafik

junaidmalik09 commented 5 years ago

did you manage to solve this?

MustafaArikan commented 5 years ago

Yes, i had mixed up the x, y dimensions for the input.

junaidmalik09 commented 4 years ago

Replacing ConvTranspose2D with Upsampling+Conv2D combo solved it for me