search

WeijingShi / Point-GNN

Point-GNN: Graph Neural Network for 3D Object Detection in a Point Cloud, CVPR 2020.
MIT License
523 stars 114 forks source link

Pedestrian & Cyclist Config #28

Closed Divadi closed 3 years ago

Divadi commented 4 years ago

Hi, I had a couple questions on the released config for ped & cyc

Here, why is the voxelization size 0.8? I remember in the paper it said 0.4 for training and 0.2 for testing. https://github.com/WeijingShi/Point-GNN/blob/2baf24f9556907f23e2e4018f1b756dac3f6c497/configs/ped_cyl_auto_T3_trainval_config#L7

Also here, it seems like it trains on the entire trainval split instead of just train. For validation on val split, we should only train on the train split, right? https://github.com/WeijingShi/Point-GNN/blob/2baf24f9556907f23e2e4018f1b756dac3f6c497/configs/ped_cyl_auto_T3_trainval_train_config#L60

WeijingShi commented 4 years ago

Hi @Divadi, Thanks for your interest.

The voxel size is "base_voxel_size" multiplied by "graph_scale". For example, base_voxel_size=0.8 and graph_scale=0.5 means a voxel of size 0.4.

https://github.com/WeijingShi/Point-GNN/blob/2baf24f9556907f23e2e4018f1b756dac3f6c497/models/graph_gen.py#L26

The ped_cyl_auto_T3_trainval_train_config trains the model for the test set. Yes, if you want to evaluate on the val split, you need to exclude the val split and use train split.

Hope this helps.