Tactile-ViewGCN: Learning Shape Descriptor from Tactile Data using Graph Convolutional Network

For humans, the sense of touch has always been crucial for precise and efficient manipulation of objects in diverse environments. However, until recently, relatively few studies have fully explored haptic feedback. In this work, we propose a novel approach that surpasses existing methods in creating shape descriptors for object classification using multiple tactile signals collected from a specialized glove. Our technique addresses the key challenge of aggregating features from multiple tactile images by considering tactile object recognition as a 3D shape recognition from multi-view problems. To do so, we introduce Tactile-ViewGCN, which hierarchically integrates tactile features and accounts for their interrelationships through a Graph Convolutional Network. When evaluated on the MIT STAG dataset, our model achieves an accuracy of 81.82%, illustrating its effectiveness in processing tactile data for object classification.