Run your deep learning-based 3D object detectors on NVIDIA Jetsons

Contents

1. Introduction
2. Environment
3. Datasets
4. Run
5. Frameworks
6. Citation
7. Acknowledgement
   
   

Introduction

This repository provides a benchmark tool for well-known deep learning-based 3D detectors on NVIDIA Jeston boards. Currently, we provide benchmarks of 12 detectors (Check (#Frameworks)!). We have tested the tool on the four Jetson series including AGX, NX, TX2, and Nano.

The work analyzes frame per second (FPS) and resource usages (CPU, GPU, RAM, Power consumption) of each detector on the Jetsons.

Clone and install requirements

1. git clone "this repository" 
2. sudo pip install -r requirements.txt

Download pre-trained weights

1. cd weights/
2. bash download_weights.sh

Environment

• Jetpack 4.4.1
• CUDA Toolkit 10.2
• Python 3.6.9
• Please check "requirements.txt" for the detailed libraries.
• The best configuration of each framework can be found in cfg folder.

Datasets

We run the benchmak using two datasets: KITTI and nuScenes. You can download the datasets from below links.

Make sure that place the datasets in 'datasets' folder.

• datasets/KITTI/*
• datasets/nuScenes/*

Dataset	Link
KITTI	link
nuScenes	link

Run

Run 'resource_anlyzer.py' in 'src/resource_analyzer' folder. You need to specify the "--model" and "--output".

$  python resource_analyzer.py --model Complex-YOLOv4 --output/C-YOLOv4  

Frameworks

Thanks for the contributors on 3D detectors. Please move to each branch for detailed instructions about source codes.

No.	Dataset	Link
1	Complex YOLOv3 w/Tiny version	link
2	Complex YOLOv4 w/Tiny version	link
3	SECOND	link
4	PointPillar	link
5	CIA-SSD	link
6	SE-SSD	link
7	PointRCNN	link
8	Part-A^2	link
9	PV-RCNN	link
10	CenterPoint	link
11	CenterPoint (TensorRT)	link

Citation

Not yet available..

@article{Soon...
}

Acknowledgement

Not yet available..