danicannt
commented
3 weeks ago Hi, We are trying a similar approach to yours, What we want is to train a PointPillars model to make it able to recognize different objects that are not among the categories included by the models.
Were you be able to share with us the steps you follow in order to reproduce your solution and to train the models? Even if they didn't work correctly
Checklist
masterbranch).My Question
I've developed a custom data loader for a custom dataset of 3 classes I have, it does not have any calibration or image files and simply contains lidar data and 3D bbox annotations. (I also made sure that the bbox annotations have the correct values by using a KITTI visualization tool on my own dataset and making sure that the bboxes are correctly placed).
I've trained several iterations of Pointpillars with different configuration inputs but I'm not getting good results even after 100-150 epochs.
Each pointcloud is around 150k points, I believe that feature extraction config part of the model is the most crucial section in the config file:
voxelize: max_num_points: 100 # Max number of point per pillar voxel_size: &vsize [0.25, 0.25, 4] max_voxels: [30000, 40000] # Number of pillars for Training,Testing
voxel_encoder: in_channels: 3 feat_channels: [64, 64] voxel_size: *vsize
scatter: in_channels: 64 output_shape: [400, 400]
yet I'm not sure how to choose the most adequate values. Any idea on how to go about it?