search

NVIDIA-AI-IOT / Lidar_AI_Solution

A project demonstrating Lidar related AI solutions, including three GPU accelerated Lidar/camera DL networks (PointPillars, CenterPoint, BEVFusion) and the related libs (cuPCL, 3D SparseConvolution, YUV2RGB, cuOSD,).
Other
1.25k stars 218 forks source link

About voxelization parameters #166

Open catziyan opened 1 year ago

catziyan commented 1 year ago

The voxelization parameters I used during training are min_range: [-27.0, 0, -5.0] max_range: [27.0, 54.0, +3.0] voxel_size: [0.075, 0.075, 0.2]

and I successfully converted the trained model to ONNX. However, when I modified the voxelization parameter in the code according to the parameters used during training, I got incorrect inference results. Surprisingly, not modifying the parameters yielded better results. Why is this?

hopef commented 1 year ago

Maybe you should double-check your training script to see what has happened.

qyang1996 commented 1 year ago

I am also meet this problem. When I use default parameters, I can obtain the correct result, but when I expand the "max_x_range", I also got wrong results.
I check the input of scn_engine. Only sparse_shape=[41, 1440, 1440] can got the correct result.

hopef commented 1 year ago

I'm sorry. I'll check it soon.

Could you provide the wrong code and model?

qyang1996 commented 1 year ago

I am using the CenterPoint model trained on my own data. training config is: voxel_generator = dict( range=[-30.0, -75.0, -5.0, 120.0, 75.0, 3.0], voxel_size=[0.075, 0.075, 0.2] ) The sparse shape will be[40, 2000, 2000].

When I modify these parameters in the common.h file, I got the incorrect 3D box. I compare the python reuslt of generateVoxels with this project, it is similar. When I manually update the sparse shape to [41, 1440, 1440] befor scn_engine_->forward(), I can get correct 3D box.

san9569 commented 3 months ago

Is there any update about this issue?

ruishanyin commented 2 months ago

I am using the CenterPoint model trained on my own data. training config is: voxel_generator = dict( range=[-30.0, -75.0, -5.0, 120.0, 75.0, 3.0], voxel_size=[0.075, 0.075, 0.2] ) The sparse shape will be[40, 2000, 2000].

When I modify these parameters in the common.h file, I got the incorrect 3D box. I compare the python reuslt of generateVoxels with this project, it is similar. When I manually update the sparse shape to [41, 1440, 1440] befor scn_engine_->forward(), I can get correct 3D box.

Did u transfer your data to nuscenes ? How?

ruishanyin commented 2 months ago

@qyang1996