Projekt- und Abschlussarbeiten

Image classification has become a central task in Machine Learning (ML), with deep learning models achieving remarkable accuracy across various datasets. Classic machine learning algorithms, not only on images but also on time series, often require expensive feature engineering steps, resulting in a demand for high computing effort and correspondingly powerful hardware. Moreover, these models often come with large memory requirements and limited interpretability. Exploring alternative approaches that are computationally efficient, lightweight, and conceptually simple is therefore of growing interest. RandOm Convolutional Kernel Transform (ROCKET) is a recent method originally developed for time-series classification that offers high efficiency through large banks of random convolutional kernels combined with simple aggregation features. Investigating whether this idea can be transferred to image data may allow insights into efficient, non-deep alternatives for image classification.

This work investigates privacy-preserving image anonymization for Ambient Assisted Living and Autonomous Driving. It focuses on realistic anonymization, which protects privacy while preserving essential image content, unlike blurring or pixelation. Using DeepPrivacy2, a fully anonymized COCO dataset is created to train and evaluate YOLOv10, analyzing effects on model performance with metrics like mAP and SSIM. The study examines how anonymization alters image characteristics, how object sizes and class co-occurrences influence detection results, and ways to improve future research. Results show challenges such as object obfuscation and re-labeling, emphasizing the need for models and datasets adapted to anonymized data.

This work addresses limitations in current methods for evaluating 6D pose estimators, which often produce low errors even for inaccurate poses, depend heavily on object-specific point clouds, and ignore false detections. Accurate evaluation is essential for robotics, augmented reality, and object manipulation, especially when analyzing performance under disturbances. The thesis proposes a novel error metric and evaluation score that work independently of object geometry and account for false detections. The score is adjustable for different evaluation scenarios. A theoretical discussion and a use case demonstrate its advantages over existing state-of-the-art metrics.