CBi-GNN

CBi-GNN: Cross-Scale Bilateral Graph Neural Network for 3D object detection

Motivation

Requirements

1. spconv
2. python >= 3.7
3. pytorch >= 1.1

Installation

1. Follow the instruction in spconv
2. build iou3d cd dets/ops/iou3d && sh install.sh
3. build pointnet2 cd dets/ops/pointnet2 && sh install.sh
4. build points_op cd dets/ops/points_op && sh install.sh

Dataset

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.

Usage

Model

• Model trained has been released on Google Drive and we will release more of different settings soon.

Test on validation

• cd excutes && python test.py ../configs/cbignn.py checkpoint_epoch_50.pth --save_to_file True --gpus=1

Kitti server

• cd excutes && python test.py ../configs/cbignn.py ../experiments/reproduce/cbignn/checkpoint.pth --save_to_file True --gpus=1 --test

Train

• cd excutes && python train.py ../configs/cbignn.py --gpus=1

Benchmark

Metrics	Easy	Moderate	Hard
recall@11	90.26	79.83	78.45
recall@40	93.36	84.35	81.15

Models

• CBi-GNN
• SECOND
• PointPillar
• PartA^2
• PV-RCNN

Datasets

• Kitti
• Waymo
• NuScenes

Based Framework

• MMCV
• PytorchLightning

Acknowledgement

This repo borrows code from the following repos:

• OpenPCDet
• SA-SSD
• 3DSSD
• SECOND
• spconv