search

OpenDriveLab / PersFormer_3DLane

[ECCV2022 Oral] Perspective Transformer on 3D Lane Detection
Apache License 2.0
416 stars 75 forks source link

The model trained with OpenLane_300 perform better than model trained with OpenLane_1000 #72

Closed yiyexy closed 1 year ago

yiyexy commented 1 year ago

A few days ago, I trained a model with OpenLane_300. I got a pretty evaluation result with this model. And then I trained another model with OpenLane_1000, but this model was nowhere near as good as the previous model.

Does anybody have the same question? Or are there any details I didn't pay attention to?

RicardLee commented 1 year ago

Hi, in contrast to OpenLane_300, the OpenLane_1000 contains more complex scene that makes the task more challenging.

yiyexy commented 1 year ago

Hi, in contrast to OpenLane_300, the OpenLane_1000 contains more complex scene that makes the task more challenging.

OK! Maybe I can understand that OpenLane_1000 contains many difficult scenes to train models. Can I trust the performance of a model trained with OpenLane_300 on the validation set?

RicardLee commented 1 year ago

We think that using OpenLane_1000 is even more convincing because the score of the baseline, PersFormer on OpenLane_1000, is open on this repo.

yiyexy commented 1 year ago

Ok! Thanks!