PhD Defense: ‘Deep Learning for Small Object Detection’
Small object detection has become increasingly relevant due to the fact that the performance of common object detectors falls significantly as objects become smaller. Many computer vision applications require the analysis of the entire set of objects in the image, including extremely small objects. Moreover, the detection of small objects allows to perceive objects at a greater distance, thus giving more time to adapt to any situation or unforeseen event.
In this PhD Thesis, the topic of small object detection has been addressed through deep learning techniques. Particularly, the work has focused on designing CNN-based architectures able to detect extremely small objects, i.e., objects under 16x16 pixels, in both still images and videos through spatio-temporal processing. Furthermore, a system aimed to automatically increase the number of instances of small objects in a given dataset has been developed, which is based on a generative adversarial network (GAN) for data augmentation.