• Super fast and accurate 3D object detection based on LiDAR
• Fast training, fast inference
• An Anchor-free approach
• No Non-Max-Suppression
• Support distributed data parallel training
• Release pre-trained models

• The technical details are described here
• The great introduction and explanation from Computer Vision and Perception for Self-Driving Cars Course Youtube link
• SFA3D is used for the second course in the Udacity Self-Driving Car Engineer Nanodegree Program: Sensor Fusion and Tracking GitHub link

Update 2020.09.06: Add ROS source code. The great work has been done by @AhmedARadwan. The implementation is here

Youtube link

The instructions for setting up a virtual environment is here.

git clone https://github.com/maudzung/SFA3D.git SFA3D
cd SFA3D/
pip install -r requirements.txt

Download the 3D KITTI detection dataset from here.

The downloaded data includes:

• Velodyne point clouds (29 GB)
• Training labels of object data set (5 MB)
• Camera calibration matrices of object data set (16 MB)
• Left color images of object data set (12 GB) (For visualization purpose only)

Please make sure that you construct the source code & dataset directories structure as below.

To visualize 3D point clouds with 3D boxes, let's execute:

cd sfa/data_process/
python kitti_dataset.py

The pre-trained model was pushed to this repo.

python test.py --gpu_idx 0 --peak_thresh 0.2

python demo_2_sides.py --gpu_idx 0 --peak_thresh 0.2

The data for the demonstration will be automatically downloaded by executing the above command.

python train.py --gpu_idx 0

• Single machine (node), multiple GPUs

python train.py --multiprocessing-distributed --world-size 1 --rank 0 --batch_size 64 --num_workers 8

• Two machines (two nodes), multiple GPUs

  • First machine

  python train.py --dist-url 'tcp://IP_OF_NODE1:FREEPORT' --multiprocessing-distributed --world-size 2 --rank 0 --batch_size 64 --num_workers 8
  

  • Second machine

  python train.py --dist-url 'tcp://IP_OF_NODE2:FREEPORT' --multiprocessing-distributed --world-size 2 --rank 1 --batch_size 64 --num_workers 8
  

• To track the training progress, go to the logs/ folder and

cd logs/<saved_fn>/tensorboard/
tensorboard --logdir=./

• Then go to http://localhost:6006/

To benchmark inference speed and collect statistics (e.g., FPS, system stats) on Jetson devices, use the provided scripts:

# For detailed stats and benchmarking
python sfa/demo_2_sides_stats.py --jclock
python sfa/demo_front_stats.py --jclock

# For live video rendering
python sfa/demo_2_sides_live.py

# Check jtop stats API
python sfa/stats.py

• The --jclock flag enables Jetson maximum performance clocks for consistent benchmarking.
• Stats scripts print detailed FPS and runtime statistics, and can output all available jtop system stats (CPU, GPU, RAM, temperature, etc.) for full hardware monitoring.
• Sample stats outputs (e.g., results/fpn_resnet_18/2sides_jclock.csv) are available in the results/ directory for reference.

Tested Jetson Environment:

• PyTorch v2.1.0
• JetPack 6.0 DP (L4T R36.2.0)
• Python 3.10
• torch-2.1.0-cp310-cp310-linux_aarch64.whl (USE_DISTRIBUTED=on)

For best performance on Jetson devices, install the official NVIDIA PyTorch wheel:

Download PyTorch for Jetson (Python 3.8, CUDA 10.2)

Install with:

pip install https://nvidia.box.com/shared/static/0h6tk4msrl9xz3evft9t0mpwwwkw7a32.whl

If you think this work is useful, please give me a star!
If you find any errors or have any suggestions, please contact me (Email: [email protected]).
Thank you!

@misc{Super-Fast-Accurate-3D-Object-Detection-PyTorch,
  author =       {Nguyen Mau Dung},
  title =        {{Super-Fast-Accurate-3D-Object-Detection-PyTorch}},
  howpublished = {\url{https://github.com/maudzung/Super-Fast-Accurate-3D-Object-Detection}},
  year =         {2020}
}

[1] CenterNet: Objects as Points paper, PyTorch Implementation
[2] RTM3D: PyTorch Implementation
[3] Libra_R-CNN: PyTorch Implementation

The YOLO-based models with the same BEV maps input:
[4] Complex-YOLO: v4, v3, v2

3D LiDAR Point pre-processing:
[5] VoxelNet: PyTorch Implementation

${ROOT}
└── checkpoints/
    ├── fpn_resnet_18/    
        ├── fpn_resnet_18_epoch_300.pth
└── dataset/    
    └── kitti/
        ├──ImageSets/
        │   ├── test.txt
        │   ├── train.txt
        │   └── val.txt
        ├── training/
        │   ├── image_2/ (left color camera)
        │   ├── calib/
        │   ├── label_2/
        │   └── velodyne/
        └── testing/  
        │   ├── image_2/ (left color camera)
        │   ├── calib/
        │   └── velodyne/
        └── classes_names.txt
└── sfa/
    ├── config/
    │   ├── train_config.py
    │   └── kitti_config.py
    ├── data_process/
    │   ├── kitti_dataloader.py
    │   ├── kitti_dataset.py
    │   └── kitti_data_utils.py
    ├── models/
    │   ├── fpn_resnet.py
    │   ├── resnet.py
    │   └── model_utils.py
    └── utils/
    │   ├── demo_utils.py
    │   ├── evaluation_utils.py
    │   ├── logger.py
    │   ├── misc.py
    │   ├── torch_utils.py
    │   ├── train_utils.py
    │   └── visualization_utils.py
    ├── demo_2_sides.py
    ├── demo_front.py
    ├── test.py
    └── train.py
├── README.md 
└── requirements.txt