Multi-sensor data augmentation for robust sensing

Abstract

Data augmentation is a crucial technique for effectively learning deep models and for improving their generalization. It has shown remarkable performance gains on complex sets of problems, such as object detection and image classification. However, for sensor (time-series) data, its potential is not thoroughly explored even though the acquisition of large annotated sensor datasets is prohibitively expensive and challenging in real-life. In this work, we propose Sensor Augment - a generalized framework for automatically discovering data-specific augmentation strategies with black-box optimization search algorithms. Our approach makes use of the user-defined transformations to discover an optimal combination of the operations that can be used to train deep networks for a wide variety of tasks. Besides, we propose several augmentation operations that can be used to generate synthetic data and enrich the search space while harnessing existing functions. We show the efficacy of learned augmentation strategies on 7 multi-sensor datasets for 4 complex tasks. In our experiments, we see a substantial performance gain ranging from 1.5 to 10 F-score points over the baseline. We also show that the strategies can be learned from smaller subsets, and they can transfer well between related datasets.