Year: 2025
Type: article
Abstract: Electro Optical (EO) and Infrared (IR) imagery provide complementary strengths for aerial object detection under diverse illumination and environmental conditions.This paper presents a reproducible, -end-to-end-YOLOv8 baseline for the ETAAV EO-IR Object Detection Challenge (Kaggle).We audit and clean labels to fix nonnormalized bounding boxes that affect evaluation and train a unified model on combined EO+IR data.The dataset is preprocessed with standard augmentation techniques and optimized hyperparameters to enhance the robustness of detection across modalities.Our experiments highlight the benefits of leveraging-both EO and IR data to handle challenging scenarios such as lowlight conditions and small object detection.Our best single-model result achieves a global F1 score of 0.4059 on the validation split, establishing-a transparent, lightweight foundation for future research on multi-modal fusion, domain adaptation, and advanced ensembling.This approach demonstrates how a simple, yet effective pipeline can be extended to improve performance with minimal computational overhead.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4413179680