Tested on Debian 9.9, Cuda: 10.0, Python: 3.6, Pytorch: 1.2.0 with Anaconda

git clone --recursive https://github.com/ziliHarvey/smart-annotation-pointrcnn.git
cd app/PointCNN/
sh build_and_install.sh

If you are using anaconda, install the environment dependencies by using environment.yaml

Testing file is at app/rl_gan/test_rl_module.py

Works on app/test_dataset/0_drive_0064_sync/sample/argoverse/lidar

cd app
python app.py

And open your browser and go to http://0.0.0.0:7772.

Testing

• Draw an approximate bounding box in one-click by pressing "a" and clicking anywhere near object
• Tick over option "Point Completion" on the left button panel to get extra points to complete the pointcloud
• Tick over option "Shape Completion" to complete the points and get a shape of object using Convex-hulling

Debugging

• Use pdb
• Use visdom provided to display output plots of completion

• Adding PointCNN Segmentation model
• Adding PointRCNN Box regresssion backend
• One-click box fitting
• Segmented object points display
• Incorporate RL-GAN-Net to do point-completion (Inference) (Training using Car Shapenet-models)
• Upgrade Encode decoder to PointCompletion Network to get robust Point Completion
• Display Shape completion with Convex hulling
• Adding Kalman filter tracking
• Use Deep learning model to do direct shape completion with image & Pointcloud