Department of Computer Engineering

The domain of object detection and classification, which is a subset of machine learning, has taken significant strides in various fields. Recently, it has shifted its focus towards achieving enhanced detection performance for objects with varying orientations. Traditional methods typically use rectangular anchor values to detect objects with (x, y, width, height) representation denoting the top-left point and width and height values. However, a new paradigm has emerged where objects are represented using (x, y, width, height, angle) parameters, encompassing center point coordinates, width, height, and rotation angle values. This novel representation aims to achieve more precise and accurate object detection. Motivations behind detecting rotated objects include the ability to discern such objects from the background more effectively, leading to improved class detection accuracy. This, in turn, facilitates sharper differentiation of object positions and accommodates the diverse ways objects can be encountered in the real world. The primary focus of this thesis is to compare and analyze models that detect rotated objects. By examining various scenarios, we ascertain their strengths and weaknesses, identifying situations in which each approach excels. Our observations are based on a comparison of one-stage, two-stage, and Transformer architectures using the DOTA-v1.0 and HRSC2016 datasets. We highlight the achievements of these models across different angles and scales.