search

lucastabelini / LaneATT

Code for the paper entitled "Keep your Eyes on the Lane: Real-time Attention-guided Lane Detection" (CVPR 2021)
https://openaccess.thecvf.com/content/CVPR2021/html/Tabelini_Keep_Your_Eyes_on_the_Lane_Real-Time_Attention-Guided_Lane_Detection_CVPR_2021_paper.html
MIT License
634 stars 167 forks source link

Bad results on kitti/custom datasets. How to improve them? #17

Closed Rajat-Mehta closed 3 years ago

Rajat-Mehta commented 3 years ago

I tried to use your pretrained model laneatt_r34_tusimple and do inference on the kitti object detection dataset (374 x 1220). As expected the results seem to be terribly bad. Capture

My question is: what are the steps that I need to follow in order to improve lane detection results on kitti or my own custom dataset?

Thanks in advance.

lucastabelini commented 3 years ago

You'll have to train on KITTI/your custom dataset. For that, you can either transform the annotations to a format already used in the code (e.g., CULane's format) or implement a new loader (from the LaneDatasetLoader abstract class). If you have issues with anything in the process tell me and I'll add a section in the README.md for that.