nuscenes_dbinfos_10sweeps_withvelo_global.npy

[Da1symeeting1]

Thanks for the great work. When I train the model(nuscens), it appears that nuscenes_dbinfos_10sweeps_withvelo_global.npy doesn't exist. And I don't know how can i creat the file? I follow the OpenPcdet dataset creation, and I got nuscenes_dbinfos_10sweeps_withvelo.pkl but no nuscenes_dbinfos_10sweeps_withvelo_global.npy.

[Say2L]

Thanks for the great work. When I train the model(nuscens), it appears that nuscenes_dbinfos_10sweeps_withvelo_global.npy doesn't exist. And I don't know how can i creat the file? I follow the OpenPcdet dataset creation, and I got nuscenes_dbinfos_10sweeps_withvelo.pkl but no nuscenes_dbinfos_10sweeps_withvelo_global.npy.

The creation of the nuscenes_dbinfos_10sweeps_withvelo_global.npy is the same as that in the waymo dataset. ( the file can save some data loading time). You can also set the USE_SHARED_MEMORY=False.

[Da1symeeting1]

I have tired replace bevfusion's Lidar branch (using voxelnet) with fshnet but it is wired that the result goes worse.
mAP: 0.6652
NDS: 0.7050
it is the result of the fusion.

[Say2L]

I have tired replace bevfusion's Lidar branch (using voxelnet) with fshnet but it is wired that the result goes worse.

We didn't integrate our method into a multimodal fusion framework. The limitations of fusion methods may differ from fully sparse detectors. Therefore, our model may not adapt well to the BEVFusion framework. And our model is not specifically designed for the nuScenes benchmark (mainly for the Waymo Open), so it may not outperform current state-of-the-art methods on nuScenes (the metrics are different across different benchmarks).