The Power of Augmentations in IR Object Detection

Recent advances in deep learning have significantly improved object detection models in various domains. However, infrared (IR) imagery introduces several challenges to these methods primarily due to limited annotated data, insufficient diversity and IR spectrum characteristics. This paper investigates the potential of data augmentation techniques for IR-based object detection. Through comprehensive experimental analysis, proposed augmentation strategy shows promising results in the anti-UAV challenge dataset. Moreover, it is independent of neural network architectures for object detection thus can be extended for various tasks easily. Furthermore, proposed approach compensates the lack of data and provides diversity in IR spectrum for training. This enables robust and reliable object detection under various IR responses depending on environmental heat and humidity distribution. We believe that IR focused augmentation presented in this paper will play a key role to enhance and improve IR applications and will be a default procedure in training of neural network models.