I have read a couple of papers on semantic segmentation and ran this github code (which was trained on Cityscapes) against a KITTI sample image and it did pretty well (as seen below).

enter image description here

I get that classification at the pixel level is very important. The problem I am having is now that I have classified each pixel, how can I use that to make decisions in autonomous drivings.

For example: the cityscapes dataset overview defines 30 classes, traffic lights and signs being in the object category.

So now that we have identified which pixels are a traffic light, we need to still get their indices so we can run that ROI through a traffic sign classification network so we can know if it's a red light, or yellow, a stop sign, etc.

Isn't a YOLOv3 model then essentially doing the same thing by putting bounding boxes around objects and doing it much cheaper and faster (computationally)?

Moreover, in the above image, there are three cars labeled in blue. So getting their indices does not actually tell me where in the image EACH car is, just where in the image cars exist.

I do see it being very useful for derivable space estimation, but is that it??

  • $\begingroup$ semantic segmentation is most important for finding the road vs. non-road regions, which is hard to get from yolo or many other approaches. There are some other things which can be better solved by semantic segmentation. AND if you combine a good semantic segmentation with an object detection, you can estimate the actual pixels of the object instead of boxes. But you could use instance segmentation for that, too (even better). $\endgroup$ – Micka Dec 26 '19 at 21:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.