Interactive labeling supports manual image labeling by presenting system predictions for users to fix errors. However, existing labeling methods do not effectively consider image difficulty, which may affect system predic-tions and user labeling. We introduce ConfLabeling, a confidence-based la-beling interface that represents image difficulties as user and system confi-dence. This interface allows users to give a confidence score to each label assignment (user confidence), and our system visualizes the results of pre-dictions with confidence levels (system confidence). We expect user confi-dence to improve system prediction, and system confidence would help us-ers quickly and correctly identify the images that need to be inspected. We conducted a user study to compare our proposed confidence-based interface with a conventional non-confidence interface in an interactive image label-ing task of varying difficulty. The results indicate that the proposed confi-dence-based interface achieved higher classification accuracy than a non-confidence interface when the image was not too difficult.
|
Chia-Ming Chang, Yi He, Xi Yang, Haoran Xie, and Takeo Igarashi. 2022. DualLabel: Secondary Labels for Challenging Image Annotation. The 48th International Conference on Graphics Interface and Human-Computer Interaction (Gl 2022), Montreal, QC, Canada, 17-19 May 2022 [PDF]
C. M. Chang, C. H. Lee, and T. Igarashi. 2021. Spatial Labeling: Leveraging Spatial Layout for Improving Label Quality in Non-Expert Image Annotation. In CHI Conference on Human Factors in Computing Systems (CHI ’21), Yokohama, Japan. May 8–13, 2021 [PDF]
C. M. Chang, S. D. Mishra and T. Igarashi, 2019, A Hierarchical Task Assignment for Manual Image Labeling. The IEEE Symposium on Visual Languages & Human-Centric Computing (VL/HCC 2019), Memphis, Tennessee, US, 14-18 October 2019 [PDF]
|
