lab-sun/InconSeg ? reverse-engineered prompt

Build me a working version of this project so I can run the road obstacle segmentation demo and training from one place.

I want the app to take in paired RGB plus depth or RGB plus disparity images and produce segmentation results for positive and negative road obstacles, like the examples in the README. Please make it easy to launch the demo and training scripts, save outputs in a clear results folder, and include a simple way to view training progress in TensorBoard.

If there are missing dataset or weight files, set up the code so it fails gracefully with a helpful message and points me to where those files should go. Use the repo’s existing Docker setup if that helps, and if anything needs current documentation or small fixes, look it up online and make the setup reliable.