Viewer-centred Completion Network (VCN)

This is the standalone code for training of the viewer-centred completion network (VCN) in SEE-VCN.

VCN is a PointNet based model that can complete the point clouds of object as captured in the wild by a lidar sensor. The coordinates of such objects are relative to the view-point of the sensor frame, which we call viewer-centred coordinates. Given an object's points, we can estimate it's pose and complete the surface without requiring pre-canonicalization like other point cloud completion methods. VCN runs at 0.32ms/car.

Install

pip install -e . --user

Datasets

• VC-ShapeNet [download]: Viewer-centred surface car dataset. Cars were positioned in viewer-centred frame using waymo labels and raycasted to obtain occluded cars in realistic scenes.
• Lidar test set [download]: We randomly select 5000 cars from KITTI, nuScenes and Waymo each for evaluation

Usage

We provide a Demo notebook with some demo data for quickstart.

Training

# Use DistributedDataParallel (DDP)
bash ./scripts/dist_train.sh <NUM_GPU> <port> \
    --config <config> \
    --exp_name <name> \
    [--resume] \
    [--start_ckpts <path>] \
    [--val_freq <int>]
    
# or just use DataParallel (DP)
bash ./scripts/train.sh <GPUIDS> \
    --config <config> \
    --exp_name <name> \
    [--resume] \
    [--start_ckpts <path>] \
    [--val_freq <int>]

For example:

# Train a model with 2 gpus
CUDA_VISIBLE_DEVICES=0,1 bash ./scripts/dist_train.sh 2 13232 \
    --config ./cfgs/VCN_models/VCN_VC.yaml \
    --exp_name exp01
    
# Resume model training
CUDA_VISIBLE_DEVICES=0,1 bash ./scripts/dist_train.sh 2 13232 \
    --config ./cfgs/VCN_models/VCN_VC.yaml \
    --exp_name exp01 --resume

Testing

bash ./scripts/test.sh 0 \
    --ckpts ./model_zoo/VCN_VC.pth \
    --config ./cfgs/VCN_models/VCN_VC.yaml \
    --exp_name exp01

Acknowledgements

Our code is built on the repository of PoinTr.