Repository files navigation

Hi! I forked repository fromultralytics version 7 to work on my undergraduate research project on KAIST Multispectral Pedestrian Dataset. I tried to apply multispectral images by merging RGB-based images and themal-based images. Please cite my work if you use my repository for your own project. For issues regarding the yolov3, please check theultralytics yolo v3.

I stop working on this project as I finished my bachelor degree. If you have any questions, I would like to answer to the best of my knowledge.

Python 3.7 or later with allrequirements.txt dependencies installed, includingtorch >= 1.5. To install run:

$ git clone https://github.com/jas-nat/PyTorch-YOLO-V3-KAIST.git$cd PyTorch-YOLO-V3-KAIST/

Installing by typing$ sudo pip3 install -r requirements.txt or$ python -m pip install -r requirements.txtI also suggest to use virtual environments, such aspyenv, docker, etc. to manage the libraries neatly.

YOLOv3 label formatUselabel_transform/kaist2coco-label-trans.py to transform label format fromKAIST Multispectral Pedestrian Dataset into YOLOv3 label format.There is also a histogram plot to draw how many people are there in the datasets.

Training configurations

Configure.cfg and.data file inconfig (See examples)-channels: 4 for multispectral, 3 for RGB, and 1 for infrared only-location of training files and validations

If you want to use pre-defined.cfg files, you can choose based on the followings

• 4 channels / multispectralyolov3-spp-1cls-4channel.cfg
• 3 channels / RGByolov3-spp-1cls.cfg
• 1 channel / infraredyolov3-spp-1cls-1channel.cfg

I also disable HSV augmentation, since it does not work for 4 channels.

Runpython3 train_kaist_multi.pySome important arguments to put afterwards:

• --weights '' to train from scratch
• --adam to use adam optimizer, the default is SGD
• --img-size to adjust the image size

Don't forget to use the correct.cfg file.Runpython3 test_kaist_multi.py.It will loadweights/best.pt. It will also produce the confidence detection results later

Don't forget to use the correct.cfg file. The image examplesRunpython3 detect_multi.py.As it is impossible to produce bounding box detections on 4 channels images, you can choose to output RGB-based or thermal-based images. You can modify the codes insidedetect_multi.py

I trained this code for 200 epochs, 4 batch size usingNVIDIA RTX 2080 Ti. It takes around 20 hours for daytime images and about 12 hours for nighttime images.

This project has been published to MDPI Journal: Sensors. Please take a look furtherhere. I humbly request you to cite our publication if you use this code as a reference.

AUTHOR = {Nataprawira, Jason and Gu, Yanlei and Goncharenko, Igor and Kamijo, Shunsuke},TITLE = {Pedestrian Detection Using Multispectral Images and a Deep Neural Network},JOURNAL = {Sensors},VOLUME = {21},YEAR = {2021},NUMBER = {7},ARTICLE-NUMBER = {2536},URL = {https://www.mdpi.com/1424-8220/21/7/2536},ISSN = {1424-8220},DOI = {10.3390/s21072536}}