[4Yin2] Interactive session 2

[4Yin2-44] Effects of Noisy Labels on Real Estate Property Image Classification

〇Hiroya Ichihara¹, Kazushi Okamoto¹, Atsushi Shibata² (1.The University of Electro-Communications, 2.Advanced Institute of Industrial Technology)

Keywords:Real Estate Property Image Classification, Noisy Labels, Annotation, Fine Tuning

In real estate industries, there are studies by using property images such as real estate property image classification, and label quality problems have been suggested recently.Each image is assigned one-to-one labels corresponding to its content such as kitchen and toilet, but there are noisy labels problems: incorrect labels; multiple labels that should be assigned; lack of suitable labels.This study defines three types of noisy labels for property images and determines the effects of the noisy labels for a property image classification task.In this study, sampled images from LIFULL HOME’S dataset are annotated in terms of the three noise types by multiple annotators.In addition, a property image classification task is performed to compare classification accuracy of the models trained by the images with and without the noisy labels.The experimental results suggest that the accuracy improves for removing images with incorrect labels and decreases for removing images that should be given multiple labels.