Investigations on the Effectiveness of Tools for Optimizing Object Detection Training Performance

Abstract

Convolutional Neural Networks (CNNs) have emerged as a pivotal technology in computer vision, particularly excelling in object detection tasks. Their adeptness at feature extraction and recognition has elevated them to the best technology for that purpose. Nevertheless, training a CNN-based object detector that performs well takes work. Often, every new scene, changed weather conditions, and other factors hamper the recognition capability of a CNN-based object detector. Thus, in practice, a re-training is necessary for every novel measurement campaign. Various tools have been devised to enhance the performance of CNN-based object detectors. Techniques like data augmentations facilitate the augmentation of training images, introducing diversity. Dropout regularization prevents overfitting, while hard example mining focuses on identifying the most challenging instances to refine the model. This thesis delves into an investigation of the efficacy of these tools in improving learning efficiency and enhancing network training. Furthermore, the research attempts to emulate the typical scenario developers face when applying a model to a new context, particularly on a customized dataset characterized by a scarcity of images and significantly different class frequencies. The focal model architecture in this study is YOLOv7, and the dataset comprises a limited collection of urban scene images with diverse perspectives. The outcomes of this research culminated in the development of multiple models, the best of which achieved a score of 80.3% on mAP@.5 and 50.7% on mAP@.5:.95, with an increase over the base model of 15.1% after training with a dataset with only 174 training images.