Skip to content

Using CSAIL vision model, a git to train your own model with your classes and the dataset you want

Notifications You must be signed in to change notification settings

Raffael1998/Road-semantic-segmentation

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

13 Commits
 
 
 
 
 
 
 
 

Repository files navigation

Road semantic segmentation

This Git teaches you how to train the CSAIL vision model with the dataset you want (see https://github.com/CSAILVision/semantic-segmentation-pytorch). The CSAIL git offers a model based on a 150 classes dataset (ADE20k dataset) = interior and outdoor / urban landscapes. If you want to train a more specific model with less classes :

The steps :

  • import the CSAIL git
  • decide on a number of classes and find (or create) a dataset that fits your needs (colormap images and reference images). The size of the images doesn't matter but I recommand not too large images
  • if not done already, you can divide your dataset to have training data and validation data
  • convert the colormap images to greyscale images using greyscale.py. This will convert the colors into greyscale from 0 to n (n = total number of classes)
  • create a .odgt file that contains the path of your greyscale images and reference images using createodgt.py
  • go to /config folder and change the training and validation path (validation data is not needed to complete a full training), the number of classes and other hyperparameters for the training if you wish -then follow the CSAIL vision git instruction to launch a training. For example python train.py --gpus GPUS --cfg config/ade20k-resnet50dilated-ppm_deepsup.yaml. Make sure you have 1 or several GPUs available, and that they are identified by pytorch.

Databases used in my case : Berkley Deep Drive and Cityscape. I kept 19 classes from Cityscape by editing all the picture and putting a black color to every color I didn't want.

About

Using CSAIL vision model, a git to train your own model with your classes and the dataset you want

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages