search

OpenDriveLab / OpenLane-V2

[NeurIPS 2023 Track Datasets and Benchmarks] OpenLane-V2: The First Perception and Reasoning Benchmark for Road Driving
https://proceedings.neurips.cc/paper_files/paper/2023/hash/3c0a4c8c236144f1b99b7e1531debe9c-Abstract-Datasets_and_Benchmarks.html
Apache License 2.0
541 stars 65 forks source link

custom dataset #88

Closed lucianzhong closed 8 months ago

lucianzhong commented 8 months ago

Hi, thanks for this great work.

By given custom dataset's camera intrinsic and extrinsic, is it possible to inference directly without training the model?

Best regards

faikit commented 8 months ago

Please consider the domain gap between training and inference data.

sephyli commented 8 months ago

Hi, if I understand correctly, you are interested in performing zero-shot transfer with models trained on OpenLane-V2. Similar to all 3D AD datasets, the main challenge of this transfer lies in the overfitting to camera parameters within BEV Encoder. There have been studies exploring this transfer, which you can further investigate.

Additionally, for the task of scene understanding, road structure significantly differ among different countries and regions, posing another major obstacle for zero-shot transfer.

lucianzhong commented 8 months ago

Thanks for the replay. Yes, zero-shot transfer. I am going to learn more about it.

lucianzhong commented 8 months ago

@faikit Thanks for the replay. domain gap, such as different data scenes and FOV of the camera