I am doing a road segmentation task for high resolution images. I have three different data sets:

  1. Around 100 with extra high resolution with the ground truth.
  2. Around 500 images with slightly worse resolution and groundtruth we extracted ourself, making it a little bit worse as well.
  3. About 2000 images with very bad resolution together with ground truth.

The model will be a CNN and tested on test set which is identical to set nbr 1.

How does one go about this problem? How should the learning rate be set and how should the training images be "shuffled" for the classifier to make optimal predictions for test data like set nbr 1?


1 Answer 1


Could you downsample all of the images to the same resolution and then feed it into a CNN? You can try to create more images as well if you use data augmentation. For road segmentation, do you mean labeled data detecting lane lines? Data augmentation could be flipping all the images horizontally, or other techniques. I think you could also try starting with a pre-trained model using transfer learning. These are just my thoughts, I'm sure some others know far more than me and could give much better answers.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.