Visual 3D Detection Package:
This repo aims to provide flexible and reproducible visual 3D detection on KITTI dataset. We expect scripts starting from the current directory, and treat ./visualDet3D as a package that we could modify and test directly instead of a library. Several useful scripts are provided in the main directory for easy usage.
We believe that visual tasks are interconnected, so we make this library extensible to more experiments. The package uses registry to register datasets, models, processing functions and more allowing easy inserting of new tasks/models while not interfere with the existing ones.
Related Paper:
This repo contains the official implementation of 2021 RAL paper Ground-aware Monocular 3D Object Detection for Autonomous Driving. Arxiv Page. Pretrained model can be found at release pages.
@ARTICLE{9327478,
author={Y. {Liu} and Y. {Yuan} and M. {Liu}},
journal={IEEE Robotics and Automation Letters},
title={Ground-aware Monocular 3D Object Detection for Autonomous Driving},
year={2021},
doi={10.1109/LRA.2021.3052442}}
Key Features
- SOTA Performance State of the art result on visual 3D detection.
- Modular Design Modular design for dataset, network and running pipelines.
- Support Various Task Compatible with the training and testing of mono/stereo 3D detection and depth prediction.
- Distributed & Single GPU Support training with multiple GPUs.
- Installation-Free Setup The setup process only build operations and does not require installation to keep the environment clean.
- Global Path-based IMDB Do not need data placed inside the folder, convienient for managing data and code separately.
We provide start-up solutions for Mono3D, Depth Predictions and more (until further publication).
Reference: this repo borrows codes and ideas from retinanet, mmdetection, M3D-RPN, DORN, EdgeNets, det3
Setup
Environment setup.
pip3 install -r requirement.txt
or manually check dependencies.
# build ops (deform convs), We will not install operations into the system environment
./make.sh
Start Training
Please check the corresponding task: Mono3D, Depth Predictions. More demo will be available through contributions and further paper submission.
Config and Path setup.
Please modify the path and other parameters in config/*.py. config/*_example files are templates.
Notice: *_examples are NOT utilized by the code and *.py under /config is ignored by .gitignore.
The content of the selected config file will be recorded in tensorboard at the beginning of training.
important paths to modify in config :
- cfg.path.data_path: Path to KITTI training data. We expect calib, image_2, image_3, label_2 being the subfolder (directly unzipping the downloaded zips will be fine)
- cfg.path.test_path: Path to KITTI testing data. We expect calib, image_2 being the subfolder.
- cfg.path.visualDet3D_path: Path to the "visualDet3D" directorty of the current repo
- cfg.path.project_path: Path to the workdirs of the projects (will have temp_outputs, log, checkpoints)
Please check the template's comments and other comments in codes to fully exploit the repo.
Further Info and Bug Issues
- Open issues on the repo if you meet troubles or find a bug or have some suggestions.
- Email to [email protected]
51 Nov 17, 2022
308 Jan 4, 2023
114 Dec 29, 2022
14 Dec 31, 2022
71 Nov 29, 2022
2 Dec 23, 2021
107 Dec 20, 2022
10 Jan 1, 2023
96 Dec 29, 2022
695 Jan 5, 2023
148 Dec 28, 2022
4 Sep 4, 2021
101 Dec 5, 2022
24 Dec 12, 2022
11 Nov 25, 2022
6 Oct 27, 2022
76 Jan 2, 2023
368 Dec 6, 2022