search

TimoSaemann / caffe-segnet-cudnn5

This repository was a fork of BVLC/caffe and includes the upsample, bn, dense_image_data and softmax_with_loss (with class weighting) layers of caffe-segnet (https://github.com/alexgkendall/caffe-segnet) to run SegNet with cuDNN version 5.
Other
176 stars 127 forks source link

accuracy decreases with weights finetuned on cityscapes #11

Closed ghost closed 7 years ago

ghost commented 7 years ago

Hy @TimoSaemann, I am very new to SegNet as well as Caffe. I tried to use the weights of the webdemo of SegNet as well as the weights that you finetuned on Cityscapes. My results looked like this: segnet Webdemo segnet citiyscapes finetuned on Cityscapes

Do you know why in this test the classes are less distinct with the weights finetuned on Cityscapes? I expected it to be the other way around. Did I get some settings wrong or does the input video have an impact or maybe does it work better with videos from inside a city? I used one from a highway.

I used @alexgkendall 's caffe-segnet and followed the tutorial. As a model filed I used the segnet_model_driving_webdemo.prototxt and just changed the weights.

Operating system: Ubuntu 14.04 Compiler: gcc 4.8.4 CPU_ONLY BLAS: ATLAS Python: 2.7

TimoSaemann commented 7 years ago

Hi @Jennifer-Mack I don't know what happend in your case, but as you pointed out, the weights finetuned on Cityscapes should lead to better results as the webdemo model.