ONE PROBLEM in the evaluation method for CULane in ERFNet-CULane-PyTorch

[yoga-0125]

Hi, I have one question regarding to the CULane evaluation method. My thought may be not correct, but please read it and give me some suggestions if possible, thank you.

• In CULane, different images may have different lane end point labelling (eg. 280, 410 ...)
• When extracting lane points from prob maps, the number of points determine the end point of lanes (eg. m=15 points result in the end points of lanes are 310)
• Predicted lane (310) can be shorter than ground truth (280) ["estimation is not enough"] or longer than ground truth (410) ["over estimation"]
• As the attached figure shown, I thought lane fitting is evaluated with almost same length, but here, different length of prediction and ground truth also affects on false positive and false negative

[cardwing]

That's true. Evaluating the lanes with the same length is wrong. You can have a look at the codes to learn how it performs evaluation.