Projekt- und Abschlussarbeiten

Camera-based systems in Ambient Assisted Living (AAL) and Autonomous Driving (AD) require careful handling of privacy-sensitive image data. The ideal way to prevent data misuse is to anonymize data right after perception and before processing. Non-realistic anonymization methods (blur, pixelation) suffice, but remove essential information needed by subsequent algorithms. Realistic anonymization, on the other hand, promises to preserve vital information, by generation of natural-like replacements. Recent studies investigate the performance on such data but do not examine the underlying causes of the observed impacts. For that reason, this study aims to establish a systematic approach to analyze anonymization methods and their effects on model training and performance, through a quantitative review of the challenges and changes introduced by anonymization.
By using the state-of-the-art toolbox DeepPrivacy2, we generate a realistic full-body anonymized COCO dataset and use it to train and evaluate YOLOv10 on object detection. In addition to classic metrics (mAP, AP), the Structural Similarity Index Measure (SSIM) is utilized to assess the impact of anonymization on images or classes. To gain insights on the influences of anonymization on computer vision, we conduct experiments focusing on factors like object size, as well as co-occurrence frequency with the anonymized class ‘person’. Furthermore, novel findings on the robustness of model sizes and the processing of anonymized images within the model are presented.
Training and evaluation with anonymized data pose challenges like object obfuscation and re-labeling. Results indicate that future research must adapt models to anonymized data, improve realistic anonymization generation, and provide datasets suited for research in anonymization. This will help establish life-changing technologies like AAL and AD and narrow the gap between privacy and the information demands of computer vision.

Current state-of-the-art methods for evaluating 6 degrees of freedom (6D) pose estimators have several significant limitations. Existing error metrics often yield near-zero errors even for inaccurate pose estimations and are highly dependent on the object point cloud used, leading to inconsistent results across different objects. Moreover, these metrics fail to account for false detections. Accurate evaluation of pose estimators is crucial for applications in robotics, augmented reality, and object manipulation, where reliable performance is essential. Evaluation is especially critical when analysing 6D pose estimators under disturbance, to gain insight on how the disturbances affect the pose estimator. This thesis introduces a novel error metric and evaluation score that can assess poses independently of the specific object and incorporate false detections. The proposed score is adjustable for various evaluation scenarios. A theoretical discussion, along with a use case analysing a 6D pose estimator under disturbances, demonstrates the advantages of the new evaluation method  compared to existing state-of-the-art approaches.