Accelerating learned descriptor generation for visual localization

• Main Author: Liu, Woon Kit
• Other Authors: Lam Siew Kei
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science
• Online Access: https://hdl.handle.net/10356/175279
• Institution: Nanyang Technological University

Visual SLAM systems use traditional feature extractor to retrieve features, a pair consisting of a keypoint and descriptor, from images. These features can then be matched to estimate the camera pose. However, these traditional feature extractors are surpassed by newer deep learning-based feature extractor in the presence of imaging noise, illumination, or viewpoint changes. However, such AI models may suffer performance issues when deployed to embedded devices, which prioritises low-powered consumption. This report investigates the potential of deep learning accelerator libraries to accelerate feature extractor models for application in visual SLAM systems, particularly on embedded devices. TensorRT, is such a library that this can help achieve a significant speedup compared to traditional feature extraction methods.