search

rasd3 / TODA

[IEEE T-IV] This is the official implementation of Semi-Supervised Domain Adaptation Using Target-Oriented Domain Augmentation for 3D Object Detection
Apache License 2.0
7 stars 0 forks source link

Where is the waymo data of modes/16^? #2

Open fdy61 opened 1 day ago

rasd3 commented 1 day ago

Hi

I used the LiDARDistill code (https://github.com/weiyithu/LiDAR-Distillation) for 16-channel downsampling. I plan to update this later, but if it’s urgent, please refer to the beam downsampling code in the getting_started.md file on the GitHub page

fdy61 commented 21 hours ago

Thanks. But Nus are of 32beams, why you use 16 instead of 32?

rasd3 commented 21 hours ago

As mentioned in the paper, the FOV of nuScenes is twice as wide as that of Waymo. Consequently, from a density perspective—meaning the number of beams per unit area—it effectively reduces to one-fourth, resulting in a similar point distribution. This is why we used a 16-channel setup.

fdy61 commented 16 hours ago

Thank you. In the code of LiDARDistill, 'ri_index' is set to 0 for only the first return in convert_range_image_to_point_cloud( https://github.com/weiyithu/LiDAR-Distillation/blob/1c222ca89685b2625260330ba137c760a1bd0e60/pcdet/datasets/waymo/waymo_utils.py#L63), whiles other codes set the ri_index to (0, 1)(TODA or SSDA3D). Does it matters?

rasd3 commented 10 hours ago

I’m not too sure about this. I just remember that create_data is from the TODA code, and the 16^ was generated using LiDARDistill.