Machine Learning


Deep Learning

Artificial neural networks are general purpose classifiers and function regressors that provide state-of-the-art performance in image classification and feature learning. With increasing number of training data available, it has become popular to learn deep architectures of multiple layers. These deep architectures often capture powerful feature representations, learned automatically from data. We apply deep neural networks to various tasks in computer vision and robotics.



Action Recognition

Recognizing human activities from sensor data streams is the foundation of many applications like anomaly detection in eldercare, surveillance and human-robot-interaction. We study approaches for action recognition on various modalities like skeleton-sequences, inertial measurement units, Wi-Fi CSI fingerprints or motion capturing data. We proposed unified approaches for the recognition of activities of known and previously unseen activities by employing Convolutional Neural Networks on motion representations or Graph Convolutional Networks.



Object Recognition

Object detection and object recognition are fundamental computer vision problems that have been studied extensively. We are particularly interested in the detection and recognition of articulated objects using color as well as depth data. We also work on part-based methods for 3D object recognition and 3D furniture recognition.



Pattern Recognition

Automated error inspection and object classification are classical examples for pattern recognition in industrial context. Together with partners from industry, we develop and integrate pattern recognition systems that feature the entire pattern recognition pipeline from sensor data preprocessing to final classification.