Title: Discovering Perceptual Attributes in a Deep Local Material Recognition Network

Abstract: Perceptual material attributes, intrinsic visual properties of materials, are being studied in parallel in computer vision and human vision. In neuroscience, perceptual attributes are shown to be an integral part of the human neural response during material recognition. In computer vision, however, they are merely an intermediate representation of materials and not integrated into the recognition process. In this paper, we show that perceptual material attributes can indeed be found inside a framework for local, patch-based, object-independent material recognition. We introduce a new CNN architecture, the material attribute-category CNN (MAC-CNN), that uses deep weak supervision to simultaneously classify materials and discover per-pixel perceptual attributes. We show that these attributes conform with past semantic material attributes and enhance recognition of novel materials. We also introduce an extensive new database for local material recognition. Our results show that the internal representation of the MAC-CNN generalizes well and agrees with human perception, which has potential implications for our understanding of human material perception as well as applications in object recognition.