search

e-lab / ENet-training

351 stars 88 forks source link

Confused about the camVid dataset used to train encoder #66

Open jay98 opened 7 years ago

jay98 commented 7 years ago

I saw that the camVid dataset downloaded from the Segnet repo has no co colors in the ground-truth, but when you get the dataset from camVid, it has ground-truth with colors in it, so I'm confused as to how exactly does it differentiate between a car and road

Thank You

codeAC29 commented 7 years ago

@jay98 we load ground-truth here. Till this point it is a grayscale image same as Segnet repo but after this we create a label tensor which is not just a single channel (grayscale) but has as many channels as number of classes.

jay98 commented 7 years ago

So the label tensor is created in the loadCamVid.lua ?

jay98 commented 7 years ago

Also @codeAC29 I'm trying to train the network on 6 classes, how would I convert the camVid images ground truths to grayscale

Thank You

codeAC29 commented 7 years ago
  1. Yes it is created in loadCamVid.lua
  2. You target label should be a tensor of dimension # classes x height x width. First channel should preferably belong to unlabeled category, given you have unlabeled category and you want to ignore it during training. Each channel will basically be a mask for each class.
jay98 commented 7 years ago

I'm new to this how exactly would I convert the ground truths? Will I have to write a script to do that?

Thank You