RefineNet: Refining Object Detectors for Autonomous Driving

• Published in: IEEE transactions on intelligent vehicles Vol. 1; no. 4; pp. 358 - 368
• Main Authors: Rajaram, Rakesh Nattoji, Ohn-Bar, Eshed, Trivedi, Mohan Manubhai
• Format: Journal Article
• Language: English
• Published: IEEE 01.12.2016
• Subjects: Autonomous driving; Autonomous vehicles; convolutional networks; fast detection; multi-perspective vision; Object detection; proposal refinement; surround behavior analysis; Vehicle detection; vehicle detection and tracking
• ISSN: 2379-8858; 2379-8904
• DOI: 10.1109/TIV.2017.2695896

Highly accurate, camera-based object detection is an essential component of autonomous navigation and assistive technologies. In particular, for on-road applications, localization quality of objects in the image plane is important for accurate distance estimation, safe trajectory prediction, and motion planning. In this paper, wemathematically formulate and study a strategy for improving object localization with a deep convolutional neural network. An iterative region-of-interest pooling framework is proposed for predicting increasingly tight object boxes and addressing limitations in current state-of-the-art deep detection models. The method is shown to significantly improve the performance on a variety of datasets, scene settings, and camera perspectives, producing high-quality object boxes at a minor additional computational expense. Specifically, the architecture achieves impressive gains in performance (up to 6% improvement in detection accuracy) at fast run-time speed (0.22 s per frame on 1242 × 375 sized images). The iterative refinement is shown to impact subsequent vision tasks, such as object tracking in the image plane and in ground plane.