Using Livox LiDAR with MMDetection3D – any experience or model suggestions?

[paureverte]

Hi everyone 👋,

I was wondering if anyone has experience using Livox LiDARs (specifically the MID360) with MMDetection3D?

As you probably know, Livox sensors use a non-repetitive scanning pattern rather than traditional ring-based LiDARs like Velodyne. This results in a different spatial distribution of points, especially per frame, which may affect how models interpret the data.

I've been thinking about trying PointPillars for a first round of training/testing, since it's voxel-based and relatively efficient. But before I proceed, I was curious if:

• Anyone has already tried Livox data with MMDetection3D?
• There are any model architectures that might handle the non-uniform pattern better?
• Any advice or gotchas to keep in mind during preprocessing, voxelization, or training?

I'll try to fine-tune with some custom data and see how it performs, but I’d appreciate any insights, feedback, or references to similar efforts.

Thanks in advance 🙏