search

Thinklab-SJTU / Bench2DriveZoo

BEVFormer, UniAD, VAD in Closed-Loop CARLA Evaluation with World Model RL Expert Think2Drive
140 stars 8 forks source link

B2D_VAD_Dataset #46

Closed Zhihua-Hua closed 2 weeks ago

Zhihua-Hua commented 2 weeks ago

In your code, the ego_lcf_feat of the B2D_VAD_Dataset has been modified, where the first two dimensions and the eighth dimension differ from those in VAD. https://github.com/Thinklab-SJTU/Bench2DriveZoo/blob/914b328f26b5534fd13b04d5fc644c8d82196d7b/mmcv/datasets/B2D_vad_dataset.py#L227 https://github.com/Thinklab-SJTU/Bench2DriveZoo/blob/914b328f26b5534fd13b04d5fc644c8d82196d7b/mmcv/datasets/B2D_vad_dataset.py#L232 https://github.com/hustvl/VAD/blob/3c2929fe36034b8881fd549394ee62f56b8ea44c/tools/data_converter/vad_nuscenes_converter.py#L508 Does it have any special meaning?

jiaxiaosong1002 commented 2 weeks ago

@Zhihua-Hua Hi, good point! It is indeed a bug. However, since the feature is fed into MLP later, it would not be a big problem and you may fix the bug and re-train the model by yourself. We do not anticipate any major influence and thus we leave it here so that the public leaderboard is stable.