Detecting the regions where a car has been damaged and the extent to which it has been damaged is a very interesting problem. It has potential applications in automatic auto insurance claims. Currently I'm trying to tackle this problem and am looking for papers/ideas for the same.

I am currently modeling the problem as a segmentation task which can classify each pixel into three classes - background, car, damaged car.

More details: I don't have a lot of data to just train a deep network on. I managed to gather around 700 damaged car images and there are multiple datasets for undamaged cars link. I've tried to generate synthetic data in Unreal using this project, but I'm afraid learning on synthetic data won't generalize to real world images.

Additionally, I found multiple startups working on the same idea but I couldn't find any info on how they're doing it link link. I also found a thesis which aims to do the same link. But it's quite old, and doesn't exactly tackle the actual problem.

I am looking for more ideas on how to formulate the problem and any previous work (like weakly supervised segmentation). Please let me know if you know of any. Thanks!

  • 2
    $\begingroup$ This question sounds too broad to me -- you're asking us how to solve an open research problem. Any community votes? This site is for focused technical questions; "any ideas that might be relevant?" is probably too open-ended. See our help center. Anyway, you probably can't determine how damaged it is just from a picture of the car: often the damage is not readily visible. (For instance, you won't be able to tell from the image whether the frame is twisted.) Also, 700 images aren't anywhere near enough to train a deep learning classifier on. $\endgroup$
    – D.W.
    Commented Apr 2, 2017 at 2:45
  • $\begingroup$ I am also looking at the same problem, for our research. Do you mind sharing your data for damaged cars with me? I understand that this is probably not the right place to ask, but it would be of great help if you agree to share the data. If you wish, we can talk more about this problem. $\endgroup$ Commented Feb 15, 2018 at 9:39