Design of tiny object detection algorithm in aerial imagery using MMdetection

• Main Author: Liu, Haoran
• Other Authors: Wen Bihan
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• Online Access: https://hdl.handle.net/10356/172237
• Institution: Nanyang Technological University

The effectiveness of object detection techniques has been greatly enhanced through the development of deep learning algorithms. Nevertheless, conventional horizontal bounding box object detection algorithms fall short in specialized scenarios such as aerial images with dense, tiny, and oriented objects. This dissertation aims to investigate enhancing the performance concerning detecting densely clustered objects of small size in aerial imagery, utilizing oriented object detection algorithms. As an extension of horizontal object detection, the oriented information of detection network excels in handling detection tasks involving many densely arranged and arbitrarily oriented targets. The research can be divided into three main contents. Initially, it envelops the significance and widespread application demand for oriented object detection, especially in the realm of aerial imagery. Following this, a comparative analysis of some of the existing benchmarks of oriented object detection algorithms is conducted on notable optical datasets like DOTA v2.0 and SAR datasets including HRSID and SSDD. This experiment provides a substantial reference for selecting foundational algorithms for subsequent optimizations. Ultimately, the research pivots towards the optimization based on R3Det, focusing on enhancing the feature fusion network, adjusting the stride of detection head, and embedding attention mechanisms within the backbone to elevate the perception capacity towards tiny objects in aerial vistas. Keywords: Tiny object detection; oriented object detection; aerial images.