Preliminary Study of a Non-Direct Generative Image Anonymization Pipeline for Anomaly Detection

With growing General Data Protection Regulation (GDPR) compliance demands for deep learning surveillance models, human anonymization is a key research area. Most studies use RGB images as input for generative models, which retain demographic features, compromising anonymization and consistency across frames. We present our initial study into a full-body anonymization pipeline for anomaly detection datasets, where the synthetic person generation never has access to the RGB pedestrian visuals. The proposed pipeline uses a combination of existing models for easier reproducibility. We use YoloV8 for object detection, ClipSeg and BiRefNet for segmentation, OpenPose for pose estimation, and an animation diffusion model. The diffusion model processes only masks and skeletal pose images, removing the problems with using sensitive data. We test on the Avenue dataset. We show that the proposed pipeline can consistently anonymize and change the demographics of detected pedestrians. We discuss the observed problems and the next steps in building a more robust second version.