3DBN

This repository contains the python implementation of 3D Backbone Network for 3D Object Detection.

NOTE-2020-5-3

The new version of thes code can be found Det3D. In the new version of 3DBN, we are using the spconv, which can make detection more efficienct and feasible for building deeper network model for higher accuracy.

Pipeline

Install

Implemented and tested on Ubuntu 16.04 with Python 3.6 and Pytorch 1.0.

Clone the repo

git clone https://github.com/Benzlxs/tDBN.git

Install Python dependencies The miniconda3 package manager package is recommended.

cd ./tDBN
pip3 install -r requirements.txt

Install Pytorch. Visiting pytorch official webpage and installing 1.0 version PyTorch according to your hardware configuration.

conda install pytorch torchvision cudatoolkit=9.0 -c pytorch

Install SparseConvNet according to its README file.
Compile the protos.

cd ./tDBN
bash protos/run_protoc.sh

Dataset

Downlaoad the KITTI and arrange files as following:

kitti_dataset
        training
                image_2
                label_2
                calib
                velodyne
                velodyne_reduced
        testing
                image_2
                calib
                velodyne
                velodyne_reduced

Split dataset and put them under the folder, ./tDBN/kitti/data_split, we provide the two set of data split, 50/50 split or 75/25 split, or you can customize your data split ratio.
Create kitti dataset infos:

cd ./tDBN
python ./scripts/create_data.py create_kitti_info_file --data_path=kitti_dataset

Create reduced point cloud:

cd ./tDBN
python ./scripts/create_data.py create_reduced_point_cloud --data_path=kitti_dataset

Create groundtruth database:

cd ./tDBN
python ./scripts/create_data.py create_groundtruth_database --data_path=kitti_dataset

Modify the directory in config file Go to the config folder and configurate the database_info_path, kitti_info_path and kitti_root_path to your path.

Training

Select your config file and output directory in train.sh, like setting config=./configs/car_tDBN_bv_2.config
Start to train:

cd ./tDBN
bash train.sh

Training results are saved in output directory, check the log.txt and 'eval_log.txt' for detailed performance.
Some training results are as following

Eval_at_125571
Car AP@0.70, 0.70, 0.70
3D  AP 87.98, 77.89, 76.35
Eval_at_160940
Car AP@0.70, 0.70, 0.70
3D  AP 88.20, 77.59, 75.58

Evaluate and inference

Select your config file and output directory in train.sh, like setting config=./configs/car_tDBN_bv_2.config
Set the model path that you want to evaluate, ckpt_path
If you want to evlaute, set test=False, if you want to generate testing result, set test=True and kitti_info_path=your_kitti_dataset_root/kitti_infos_test.pkl
Start to evaluate or inference:

cd ./tDBN
bash evaluator.sh

Testing results on KITTI benchmark.

Benchmark	Easy	Moderate	Hard
Car (Detection)	90.30 %	88.62 %	80.08 %
Car (Orientation)	89.93 %	87.95 %	79.32 %
Car (3D Detection)	83.56 %	74.64 %	66.76 %
Car (Bird's Eye View)	88.13 %	79.40 %	77.97 %

Acknowledge

Thanks to the team of Yan Yan , we have benifited a lot from their previous work SECOND.

Benzlxs / tDBN