69.9 on the CULane but misalignment with GT in some scene

cfzd / Ultra-Fast-Lane-Detection

Ultra Fast Structure-aware Deep Lane Detection (ECCV 2020)

MIT License

1.82k stars 493 forks source link

69.9 on the CULane but misalignment with GT in some scene #11

Closed wohaiyo closed 4 years ago

wohaiyo commented 4 years ago

Hello, I got a 69.9 on the CULane dataset using a single GPU card with your default settings. I visualized the lane detection results and found that the model performed well in some ideal scene. But the predicted results seriously deviate from the GT when it comes to occlusion, Vehicle lane changing, ambiguous appearance. Do you have the same phenomenon? Is this a limitation of the method? backbone? or something else?

cfzd commented 4 years ago

@wohaiyo Hi, I'm not sure what's the problem you are talking about. The youtube demo shows our results on CULane.

Can you show these results?