TOWARDS EFFICIENT OBJECT DETECTION WITH DEEP LEARNING

Abstract

Object detection aims at localizing and recognizing object instances of certain classes in the input image. It is a fundamental computer vision task that enables instance-level visual perception and thus is important for many vision applications.

These algorithms are mostly developed in laboratory settings with well-curated benchmarks and modern high-performance computation devices. However, they face significant challenges in real-world scenarios like computation limitation and environmental variation. The focus of this thesis is to develop solutions addressing the mentioned challenges and achieve efficient object detection. Concretely, we take two perspectives, model efficiency and data efficiency. We propose detection knowledge distillation and feature compression technique to accelerate model inference. We establish few-shot adaptation detection and long-tail detection frameworks to achieve efficient data usage. The proposed methods are extensively validated on various experimental settings and benchmarks, we also conduct extensive ablation and analysis to better understand the effectiveness of the methods. We believe the findings and methods developed in the thesis provide insights for future research on object detection.