OpenDriveLab / ViDAR

[CVPR 2024 Highlight] Visual Point Cloud Forecasting
https://arxiv.org/abs/2312.17655
Apache License 2.0
235 stars 15 forks source link

Can Vidar produce 4D binary occupancy results during inference? #6

Closed synsin0 closed 4 months ago

synsin0 commented 4 months ago

Thanks for your great work! I have doubt about the latent rendering during inference between yours vidar and 4d-occ-forecasting. They generate query pred_pcds from a binary 4D occupancy grid. Can Vidar produce 4D binary occupancy results during inference? Or Vidar can only generate future pointclouds?

tomztyang commented 4 months ago

It depends on the supervision in my opinion. You can still follow ViDAR architecture and use occupancy ground-truth as supervision. Then, you can get the 4D binary occupancy.

I think, there is no difference between 4D occupancy and future point cloud predictions in essential. For point cloud forecasting, we just need to render those points along different rays to obtain point clouds based on the 4D occupancy volume.

Any further discussions are welcomed!

Best, Zetong