Submission instruction

[yuanpengtu]

Hi, when will the submission instruction be available?

[faikit]

Hi, it will be make avaliable before the submission period open.
We also provide an option in baseline model to format a pickle file for submission.

[yuanpengtu]

Hi，I wonder what hardware is used to train the baseline model， i.e., the graphics card and the used number.

[faikit]

It takes about 20 hours on 8 A100. And it could also fit into a 3090.

As asked from other platforms, to reduce memory consumption, you change the modules of the network, e.g., the DETR head to deformable DETR.
The baseline is designed for easy modification and extension. So it is recommended to get familiar with the tasks through our baseline, and then implement your awesome thoughts to solve the task.

[yuanpengtu]

It seems that I cannot reproduce the result of the baseline with 8 V100, where the batch size can only be 1 for each card. Is the batch size influence the performance? Actually I have adjusted the learning rate according to the batch size.

[yuanpengtu]

DET_l: 0.0033836588263511658, DET_t: 0.07977674901485443, TOP_ll: 0.0000, TOP_lt: 0.0000, OpenLane-V2 Score: 0.0222, F-Score for 3D Lane: 0.0818. This is the reproduced performance

[faikit]

Hi, the model is trained on some flawed data and the scores are reported on the test set.
But as mentioned in #3 that the baseline is provided for quick modification. So I think it is by no means crucial to align the scores.

Your scores are decent on this basic model based on my experience.
I would recommend you start to design your model and have fun.