Multimodal Multi-User Surface Recognition With the Kernel Two-Sample Test

Machine learning and deep have been used extensively to classify physical surfaces through images time-series contact data. However, these methods rely on human expertise entail the time-consuming processes of data parameter tuning. To overcome challenges, we propose an easily implemented framework that can directly handle heterogeneous sources for classification tasks. Our data-versus-data approach automatically quantifies distinctive differences in distributions a high-dimensional space via kernel two-sample testing between two sets extracted from multimodal (e.g., images, sounds, haptic signals). We demonstrate effectiveness our technique by benchmarking against expertly engineered classifiers visual-audio-haptic surface recognition due industrial relevance, difficulty, competitive baselines this application; ablation studies confirm utility key components pipeline. As shown open-source code, achieve 97.2% accuracy standard multi-user dataset with 108 classes, outperforming state-of-the-art machine-learning algorithm 6% more difficult version task. The fact classifier obtains performance minimal processing setting reinforces powerful nature recognize complex patterns. Note Practitioners —We how apply test surface-recognition task, discuss opportunities improvement, explain use other problems similar properties. Automating could benefit both inspection robot manipulation. class similarity therefore outputs ordered list surfaces. This is well suited quality assurance documentation newly received materials or manufactured parts. More generally, automated pipeline including high-frequency measurements vibrations, forces signals. circumvents process feature engineering, experts non-experts it high-accuracy classification. It particularly appealing new without existing models heuristics. In addition strong theoretical properties, straightforward practice since requires only evaluations. Its transparent architecture provide fast insights into given case under different sensing combinations costly optimization. Practitioners also procedure obtain minimum data-acquisition time independent sensor recordings.