Artificial Fish-Eye Image Augmentation Approach for CNN Based Object Detection

Not always, the data needed for training neural network models is available in required amount. A clear example of this are images taken with fish-eye cameras - they a rewidely being used due to their ability to capture a large field of view, but there are not many publicly available datasets for specific tasks. In this paper, we research how to deal with data scarcity in object detection by augmenting images using different fisheye camera models to obtain more training data. At first, we test and implement several fish-eye camera models to obtain artificial fish-eye images using various distortion scenarios. Then a method for recreating the bounding box labels used for training the object detection model is proposed. Using the different fish-eye camera models with various distortion scenarios and bounding box recreation several training datasets with artificial fish-eye images were created. Finally, datasets created using our proposed approach are used for training the widely used YOLOv8 object detection model, and the models were then tested on datasets with images taken from real fish-eye camera. The results clearly demonstrate that the proposed augmentation approach is viable to deal with fish-eye image scarcity, as models trained on augmented datasets showed better performance across many metrics than a model trained without augmentation.