Abstract
Direct time-of-flight (ToF) sensors measure photon transients - the distribution of returned photons across return times-which encode rich information beyond depth.We present a computational approach that leverages these transients for robust, non-contact material classification using a miniature, multiregion ToF sensor. Our approach classifies photon transient data with a fully-connected neural network to achieve a 98.4% test accuracy under varying distances and orientations. Experiments cover a wide range of material rotation angles and sensor-to-object distances so that the classification generalizes over capture conditions. To improve generalization and robustness to ambient illumination we introduce data augmentation techniques during training. Beyond material classification we contribute a Bayes factor analysis of transients that highlights statistically significant shape differences. This analysis offers evidence of material classification signatures and generally aids the interpretation of photon transients. The ToF sensor consists of a low-resolution array of regions of interest (ROIs), with each ROI capturing distinct scene points. We also develop techniques that calibrate multi-region ToF sensors to allow for the measurement of reflectance-based material signatures that complement the data-driven neural network classification. This study demonstrates practical methods to leverage photon transient data from low-cost, miniaturized ToF sensors to support new imaging modalities that extend beyond depth.