Home // MMEDIA 2019, The Eleventh International Conference on Advances in Multimedia // View article

Rapid Annotation Tool to Train Novel Concept Detectors with Active Learning

Authors:
Maaike H.T. de Boer
Henri Bouma
Maarten Kruithof
Bart Joosten

Keywords: image annotation; concept localization; deep learning; active learning

Abstract:
Annotating a large set of images, especially with bounding boxes, is a tedious task. In this paper, we propose an intuitive image annotation tool. This tool not only allows (non-expert) users to annotate images with novel concepts, but is also able to achieve acceptable performance with a smaller amount of annotated images. The tool can also propose detections on unannotated images, to provide faster annotation and insight in the performance of the system. The tool is based on an Single Shot Multi-box Detector (SSD) neural network with active learning, based on showing the images with high-confident detections first, to have a fast verification and re-training. An experiment on simulated data shows that this active learning method can achieve higher performance in a shorter expected annotation time with a small number of images (less than 500). A small experiment on user annotated data shows that the annotation tool allows faster annotation compared to without the annotation tool.

Pages: 36 to 41

Copyright: Copyright (c) IARIA, 2019

Publication date: March 24, 2019

Published in: conference

ISSN: 2308-4448

ISBN: 978-1-61208-697-2

Location: Valencia, Spain

Dates: from March 24, 2019 to March 28, 2019