Indoor Space Recognition using Deep Convolutional Neural Network: A Case\n Study at MIT Campus

Abstract

In this paper, we propose a robust and parsimonious approach using Deep\nConvolutional Neural Network (DCNN) to recognize and interpret interior space.\nDCNN has achieved incredible success in object and scene recognition. In this\nstudy we design and train a DCNN to classify a pre-zoning indoor space, and\nfrom a single phone photo to recognize the learned space features, with no need\nof additional assistive technology. We collect more than 600,000 images inside\nMIT campus buildings to train our DCNN model, and achieved 97.9% accuracy in\nvalidation dataset and 81.7% accuracy in test dataset based on spatial-scale\nfixed model. Furthermore, the recognition accuracy and spatial resolution can\nbe potentially improved through multiscale classification model. We identify\nthe discriminative image regions through Class Activating Mapping (CAM)\ntechnique, to observe the model's behavior in how to recognize space and\ninterpret it in an abstract way. By evaluating the results with\nmisclassification matrix, we investigate the visual spatial feature of interior\nspace by looking into its visual similarity and visual distinctiveness, giving\ninsights into interior design and human indoor perception and wayfinding\nresearch. The contribution of this paper is threefold. First, we propose a\nrobust and parsimonious approach for indoor navigation using DCNN. Second, we\ndemonstrate that DCNN also has a potential capability in space feature learning\nand recognition, even under severe appearance changes. Third, we introduce a\nDCNN based approach to look into the visual similarity and visual\ndistinctiveness of interior space.\n