Skip to content

Creating virtual 2D points via multi-depth projection

Notifications You must be signed in to change notification settings

SxJyJay/MDU_preprocess

Repository files navigation

multi-depth-projection

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.

nuscenes segmentation results generation

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.

About

Creating virtual 2D points via multi-depth projection

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages