lhoyer/improving_segmentation_with_selfsupervised_depth

Training with custom dataset

Closed this issue · 5 comments

gruossomonica commented

Hi,
thank you for your interesting work.
I'm interested in semi-supervised segmentation. I have a question. Can I perform training using your approach with my own dataset? I collected about 5k images with binary segmentation masks (foreground and background classes). What are the tips and steps in this case?
I look forward to your reply. Thank you.

Best Regards,

Monica

lhoyer commented

Thank you for your interest in our work. To apply our method, your dataset has to fulfill the following criteria:

  • there is access to the underlying video clips in order to train SDE,
  • some of the images from the video clips have segmentation labels, and
  • the camera intrinsics have to be known.

If that is the case, you can check out the class CityscapesLoader in loader/cityscapes_loader.py for a reference on how to implement your own data loader. It inherits from loader/sequence_segmentation_loader.py, where you can find out more about the underlying data loading logic. In order to adapt the framework to the new data loader, a search with the keyword "cityscapes" will reveal the relevant places. After that, the README.md will guide you through the training process.

  • 👍1
gruossomonica commented

Thanks for the detailed explanation.

Best Regards,

Monica

AnukritiSinghh commented

Hey @gruossomonica, were you able to run EXP 210 or 211? how did you place your dataset in the folders?
Thank you

gruossomonica commented

Hi @AnukritiSinghh! Sorry, I didn't try because I didn't have the information about the camera intrinsics and more than one camera was used to collect my dataset.

Best regards.

fjremnav commented

@gruossomonica

Have you succeeded to modify the codes for your own datasets?

Thanks,