** 3 4 .**

4ch-YOLOv8

**4ch-YOLOv8** GOP                 6    RGB + IR()  4    YOLOv8    .

RGB *IR() 4 * .

Key Modification

YOLOv8

• YOLOv8 Conv layer(conv1) 34
• conv1 , pretrained weights

Project Environment

• OS: Ubuntu 25.04
• Python: 3.10
• Pytorch:2.2.0
• CUDA:12.1
• GPU: NVIDIA TITAN Xp / A100 (Colab)

Colab Notebook

Thermal-CycleGAN Tutorial: |

Getting Started

Installation

• Clone this repo: bash git clone https://github.com/scksh/4ch-YOLOv8 cd 4ch-YOLOv8
• install libraries: bash pip install -r requirements.txt pip install torch==2.7.0 torchvision==0.22.0 torchaudio==2.7.0 --index-url https://download.pytorch.org/whl/cu126 pip install -e ../4ch-YOLOv8

Download Dataset (HuggingFace)

• You can download the RGBIR dataset directly from Hugging Face Hub using the following Python code: ```python from huggingfacehub import hfhub_download import zipfile

Download dataset

zippath = hfhubdownload( repoid="SUMMERZETT/RGBIR", filename="RpGBIR.zip", repo_type="dataset" )

with zipfile.ZipFile(zippath, 'r') as zipref: zip_ref.extractall("./datasets") ```

Download Initial weight

• To download the Initial weight for 4ch-YOLOv8 model from Hugging Face, use the following code: ```python from huggingfacehub import hfhub_download

Download pretrained model

modelpath = hfhubdownload( repoid="SUMMERZETT/YOLOv8pretrained", filename="yolov8x4chpretrained.pt", repotype="model", localdir="./model", localdirusesymlinks=False ) ```

Download Pre-trained model

• To download the pretrained 4ch-YOLOv8 model from Hugging Face, use the following code: ```python from huggingfacehub import hfhub_download

Download pretrained model

modelpath = hfhubdownload( repoid="SUMMERZETT/4ch-YOLOv8", filename="4ch-YOLOv8.pt", repotype="model", localdir="./pretrained", localdiruse_symlinks=False ) ```

Train

• Train a model: bash export KMP_DUPLICATE_LIB_OK=TRUE yolo task=detect train model=model/yolov8x_4ch_pretrained.pt data=ultralytics/cfg/datasets/RGBIR.yaml workers=2 epochs=1 batch=8 cos_lr=True

Prediction

• Predict the results: bash python predict.py --weights pretrained/4ch-YOLOv8.pt --source_rgb datasets/RGBIR/images/rgb_val --source_ir datasets/RGBIR/images/thermal --project runs/predict --name 4ch-YOLOv8_pred --imgsz 640 --conf 0.4 --iou 0.5 --device cuda

Citation

If you use this code for your research, please cite our papers. @misc{thermalcyclegan2025, title={4ch-YOLOv8: Object detection using 4channel images contain RGB+IR}, author={Cha, Hyunwoo and Do, Jihoon and Gang, Nayoon and Kim, Seunghwan and Lee, Haerin and Yoon, Youngbin}, year={2025}, howpublished={\url{https://github.com/scksh/4ch-YOLOv8}}, note={GitHub repository} }

Acknowledgments

Our code is inspired by - ultralytics