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

Why we can't perform per-class accuracy like training demo in Cityscapes Data? #85

Open nurhadiyatna opened 7 years ago

nurhadiyatna commented 7 years ago

Hi @alexgkendall, I am a new user of caffe and segnet, I currently running Segnet Training with Cityscapes data with 360x480 pixels of dimension (resized from original size). I use Titan X 12 GB, with only 1 Batch size, and run for 40000 iterations. However, I have few question:

  1. Which one of the label that we should ignore in cityscapes data?
  2. Why we can't perform per-class accuracy like in training demo with CamVid Dataset?

I deleted this part from segnet_train.prototxt:

layer { name: "accuracy" type: "Accuracy" bottom: "conv1_1_D" bottom: "label" top: "accuracy" top: "per_class_accuracy" }

and the training run well. But why?