Assessing the YOLO Series Through Empirical Analysis on the KITTI Dataset for Autonomous Driving

• Type: Article in International Conference Proceedings Book
• Date: 2020

Authors

Filipa Ramos (Author) FEUP; Alexandre Correia (Author) Other; Rosaldo J. F. Rossetti (Author) FEUP

Conference proceedings International

Title: INTELLIGENT TRANSPORT SYSTEMS

Pages: 203-218

3rd European-Alliance-for-Innovation (EAI) International Conference on Intelligent Transport Systems (INTSYS)

Braga, PORTUGAL, DEC 04-06, 2019

Indexing

ISI Web of Knowledge - 0 Citations

ISI Web of Science

Scopus - 1 Citation

Other information

Authenticus ID: P-00R-RVB

DOI: 10.1007/978-3-030-38822-5_14

Abstract (EN): Computer vision and deep learning have been widely popularised on the turn of the 21st century. On the centre of its applications we find autonomous driving. As this challenge becomes a racing platform for all companies, both directly and indirectly involved with transportation systems, it is only pertinent to evaluate exactly how some generic, state-of-the-art models can perform on datasets specifically built for autonomous driving research. With this purpose, this article aims at directly studying the evolution of the YOLO (You Only Look Once) model since its first implementation until the most recent version 3. Experiences carried out on the respected and acknowledged driving dataset and benchmark known as KITTI Vision Benchmark enable direct comparison between the newest updated version and its predecessor. Results show how the two versions of the model have a performance gap whilst being tested on the same dataset and using a similar configuration setup. YOLO version 3 shows its renewed boost in accuracy whilst dropping minimally on detection speed. Some conclusions on the applicability of models such as this to a real-world scenario are drawn so as to predict the direction of research in the area of autonomous driving.

Language: English

Type (Professor's evaluation): Scientific

No. of pages: 16