Automatic Assessment of Hoarding Clutter from Images Using Convolutional Neural Networks

Abstract

Hoarding is a mental and public health problem stemming from difficulty associated with discarding one’s possessions and resulting clutter. In the last decade, a visual method, called "Clutter Image Rating" (CIR), has been developed for the assessment of hoarding severity. It involves rating clutter in patient’s home on the CIR scale from 1 to 9 using a set of reference images. Such assessment, however, is time-consuming, subjective, and may be non-repeatable. In this paper, we propose a new automatic clutter assessment method from images, according to the CIR scale, based on deep learning. While, ideally, the goal is to perfectly classify clutter, trained professionals admit assigning CIR values within ±1. Therefore, we study two loss functions for our network: one that aims to precisely assign a CIR value and one that aims to do so within ±1. We also propose a weighted combination of these loss functions that, as a byproduct, allows us to control the CIR mean absolute error (MAE). On a recently-collected dataset, we achieved ±1 accuracy of 82% and MAE of 0.88, significantly outperforming our previous results of 60% and 1.58, respectively.