Tianwei Yin
Tianwei Yin
just a comment, it seems the output stride is important for centerpoint performance on KITTI. In current configs (nusc exps), we use 8x for voxelnet, but I found that 4x...
> By the way, can you confirm if the feature map dimensions are appropriate for the KITTI point cloud data range? Which part do you want me to confirm? I...
@robin-karlsson0 I didn't change the 3d backbone and only add a 2x upsample at the end of the 2d backbone (before detection head). This should take less memory. ``` pts_middle_encoder=dict(...
@tomztyang Is there inference time statistics on nuScenes? I want to reference this in a table.
thanks!
> can you explain what each item means? Besides, can you tell me how the voxels data generates? where is the corresponding code? metadata': this contains infos for this frame...
> Because its implementation simply traverses the point cloud step by step, generates the voxel index, and fills the point cloud's feature data into the new voxel matrix according to...
Hi, thanks for your interest. I will have a large update to the codebase in the coming weeks. And I will make sure to add some code for own data...
@huster-wugj this is a community contribution. You can find details here https://github.com/tianweiy/CenterPoint/pull/11 You can't directly use /velodyne_points currently. I will add script to support this later.
I guess the problem is solved. Reopen if you still have some issues