Department of Computer Engineering

Object Detection, one of the landmark problems in Computer Vision, is conventionally addressed by placing horizontal rectangular boxes around objects. The use of such bounding boxes, however, puts a restriction about the orientations of the objects. For elongated and oriented objects, such boxes usually include disruptive visual information coming from background or other objects, which can lead to poor detection performance for such elongated and oriented objects.

To address this issue, in Oriented Object Detection (OOD), objects are described and detected with oriented rectangular bounding boxes. The introduction of many large-scale datasets with oriented bounding box annotations, e.g. the DOTA-v1.0 dataset, has facilitated a plethora of approaches for OOD. However, to the best of our knowledge, there is no study that extensively analyzes the issues pertaining to OOD problem. In this thesis, we provide detailed analyzes using the DOTA-v1.0 dataset and show e.g. that (i) there is a severe imbalance problem regarding the distribution of objects across different orientations and scales, (ii) especially for certain object classes (e.g.small vehicle, bridge, harbor), and (iii) the skewness of these distributions leads to lower performance for orientations and scales with less samples.