cfzd / Ultra-Fast-Lane-Detection

Ultra Fast Structure-aware Deep Lane Detection (ECCV 2020)
MIT License
1.82k stars 493 forks source link

Running on own images #44

Closed AlexKaravaev closed 4 years ago

AlexKaravaev commented 4 years ago

I have managed to run on my own images, however, I can not understand why end lines are a little bit shifted. It seems to me that pretrained weights are working well, but there are maybe some postprocessing steps, that require to change something if I want to use own images, instead of culane/tusimple.

This is the images I got from running demo.py https://imgur.com/a/PdeTRC9

AlexKaravaev commented 4 years ago

Okay, I've managed to get it to work. If someone will encounter this issue, the model is sensitive to the resolution of the data it was trained on. Firstly, I was trying to feed 1640×590 images on Tusimple weights. Images are from that run.

Things are good, when you run with original resolution(Or at least scale ration needs to be the same I guess).

So this are correct resolutions for the weights