Effects of Noisy Labels on Real Estate Image Classification

DOI IR

Bibliographic Information

Other Title
  • 不動産物件画像分類におけるラベルノイズの影響

Abstract

In the real estate industry, studies using property images for tasks such as property image classification have attracted considerable interest, and recent studies have reported problems with label quality. Each image is assigned one-to-one labels according to its content such as kitchen and toilet, but there are noisy labelling problems: incorrect labels; multiple labels that should be assigned; lack of appropriate labels. This study defines three types of noisy labels for property images and determines their distributions and the effects of the noisy labels on a property image classification task. In this study, 30,800 sampled images from LIFULL HOME'S dataset were annotated for the three noise types by multiple annotators. It was confirmed that 17% of these images contained label noise. In addition, in the property image classification task using images both with and without noise labels, the results suggest that removing images with incorrect labels improved the accuracy, while removing images that should have multiple labels led to a decrease in accuracy.

Journal

Keywords

Details 詳細情報について

Report a problem

Back to top