Accelerating Active Learning Image Labeling Through Bulk Shift Recommendations
- Nowadays, the inexpensive memory space promotes an accelerating growth of stored image data. To exploit the data using supervised Machine or Deep Learning, it needs to be labeled. Manually labeling the vast amount of data is time-consuming and expensive, especially if human experts with specific domain knowledge are indispensable. Active learning addresses this shortcoming by querying the user the labels of the most informative images first. One way to obtain the ‘informativeness’ is by using uncertainty sampling as a query strategy, where the system queries those images it is most uncertain about how to classify. In this paper, we present a web-based active learning framework that helps to accelerate the labeling process. After manually labeling some images, the user gets recommendations of further candidates that could potentially be labeled equally (bulk image folder shift). We aim to explore the most efficient ‘uncertainty’ measure to improve the quality of the recommendations such that all images are sorted with a minimum number of user interactions (clicks). We conducted experiments using a manually labeled reference dataset to evaluate different combinations of classifiers and uncertainty measures. The results clearly show the effectiveness of an uncertainty sampling with bulk image shift recommendations (our novel method), which can reduce the number of required clicks to only around 20% compared to manual labeling.
Author: | Philipp Scharpf, Chi Lap Hong, Oliver DürrORCiDGND |
---|---|
DOI: | https://doi.org/10.1109/ICDMW53433.2021.00055 |
ISBN: | 978-1-6654-2427-1 |
ISBN: | 978-1-6654-2428-8 |
Parent Title (English): | International Conference on Data Mining Workshops (ICDMW), 7.-10. Dezember 2021, Auckland, New Zealand |
Publisher: | IEEE |
Document Type: | Conference Proceeding |
Language: | English |
Year of Publication: | 2021 |
Release Date: | 2022/08/02 |
Tag: | Active Learning; Computer Vision; Incremental Classification and Clustering; Image Classification; Image Recognition; Image Labeling |
First Page: | 398 |
Last Page: | 404 |
Note: | Volltextzugriff für Angehörige der Hochschule Konstanz möglich. |
Relevance: | Keine peer reviewed Publikation (Wissenschaftlicher Artikel und Aufsatz, Proceeding, Artikel in Tagungsband) |
Open Access?: | Nein |
Licence (German): | Urheberrechtlich geschützt |