Creating virtual 2D points via multi-depth projection
Setup steps please refer to MVP.
Based on MVP's framework, we further devise two virtual points projection approaches, i.e., multi-depth-projection and boundary-projection. The multi-depth-projection (50 pts, 6NN) can bring over 2.0 mAP improvements on nuScenes dataset, while boundary-projection has not been experimented. To create multi-depth projection data, you can run the following command:
python virtual_depth_mapping_multi_proj.py --info_path data/nuScenes/infos_train_10sweeps_withvelo_filter_True.pkl MODEL.WEIGHTS centernet2_checkpoint.pth
Similarly, you can also create boundary projection data by running:
python virtual_boundary_gen.py --info_path data/nuScenes/infos_train_10sweeps_withvelo_filter_True.pkl MODEL.WEIGHTS centernet2_checkpoint.pth
You can also create depth completion-based projection data (at most 10,000 pts for each instance) by running:
python virtual_depth_completion_nuscenes.py --info_path data/nuScenes/infos_train_10sweeps_withvelo_filter_True.pkl MODEL.WEIGHTS centernet2_checkpoint.pth
If you want to increase or reduce the number of pts per instance, please modify the upper bound of random selection here.
Don't forget to modify the path of saving the created data.
Generate nuscenes instance segmentation results and save them as json files.
You can use the following command to generate nuscenes segmentation results files.
python nusc_seg.py --info_path data/nuScenes/infos_train_10sweeps_withvelo_filter_True.pkl MODEL.WEIGHTS centernet2_checkpoint.pth
Don't forget to modify the relevant file reading and saving path.