CBi-GNN: Cross-Scale Bilateral Graph Neural Network for 3D object detection
- spconv
- python >= 3.7
- pytorch >= 1.1
- Follow the instruction in spconv
- build iou3d
cd dets/ops/iou3d && sh install.sh
- build pointnet2
cd dets/ops/pointnet2 && sh install.sh
- build points_op
cd dets/ops/points_op && sh install.sh
Dataset is downloaded from KITTI and orgnized as follow:
├── data
│ ├── KITTI
│ │ ├── ImageSets (txt files for splited samples list (train.txt, val.txt, test.txt))
│ │ ├── object
│ │ │ ├──training
│ │ │ ├──calib & velodyne & label_2 & image_2 & (optional: planes)
│ │ │ ├──testing
│ │ │ ├──calib & velodyne & image_2
├── lib
├── pointnet2_lib
├── tools
This dataset include 7481 samples for training and 7518 samples for online test. We also split training dataset into train and val for improving model following OpenPCDet, specifically 3769 samples for val and 3712 samples for train.
- Model trained has been released on Google Drive and we will release more of different settings soon.
cd excutes && python test.py ../configs/cbignn.py checkpoint_epoch_50.pth --save_to_file True --gpus=1
cd excutes && python test.py ../configs/cbignn.py ../experiments/reproduce/cbignn/checkpoint.pth --save_to_file True --gpus=1 --test
cd excutes && python train.py ../configs/cbignn.py --gpus=1
Metrics | Easy | Moderate | Hard |
---|---|---|---|
recall@11 | 90.26 | 79.83 | 78.45 |
recall@40 | 93.36 | 84.35 | 81.15 |
- CBi-GNN
- SECOND
- PointPillar
- PartA^2
- PV-RCNN
- Kitti
- Waymo
- NuScenes
- MMCV
- PytorchLightning
This repo borrows code from the following repos: