Fine-Tuning Segformer for Improved Lane Detection in Autonomous Vehicles

[@LearnOpenCV]

Check out our FREE Courses at OpenCV University : opencv.org/university/free-courses/

[@rahulpai3393]

The training codenis stuck at 1st epoch Epoch 1/30: 0%| | 0/1750 [00:00

[@LearnOpenCV]

The information provided is insufficient to give a solution.

[@rahulpai3393]

@@LearnOpenCV It worked out when I tried on a higher spec PC. THanks!!

[@eusandre95]

How did you get those segmented ground truth masks with the lanes completely filled? In the original dataset they provide only the lane boundaries which have "holes" inside. Did you perform some kind of pre processing to the original ground truth masks? Thanks in advance.

[@LearnOpenCV]

Good question! In some cases, you can add additional pre-processing steps in the training DataLoader to account for such imperfections in the GT masks.

[@eusandre95]

Thanks for amswering. But I didn't get if you performed this kind of processing to the original ground truth images or not. Did you actually manipulate the GT images to "fill" the holes inside the lane marking boundaries? If yes, how? I looked at the code in the blog but it doesn't seem that you do that. I would appreciate if you can share the code or at least the idea behind. Thanks