Does Image Anonymization Impact Computer Vision Training?
This addresses the trade-off between privacy compliance and data utility for computer vision developers, offering an incremental improvement by showing realistic anonymization can mitigate performance loss.
The paper investigates how image anonymization affects computer vision model training, finding that traditional anonymization significantly reduces performance, especially for full bodies, while realistic anonymization minimizes the drop, particularly for faces.
Image anonymization is widely adapted in practice to comply with privacy regulations in many regions. However, anonymization often degrades the quality of the data, reducing its utility for computer vision development. In this paper, we investigate the impact of image anonymization for training computer vision models on key computer vision tasks (detection, instance segmentation, and pose estimation). Specifically, we benchmark the recognition drop on common detection datasets, where we evaluate both traditional and realistic anonymization for faces and full bodies. Our comprehensive experiments reflect that traditional image anonymization substantially impacts final model performance, particularly when anonymizing the full body. Furthermore, we find that realistic anonymization can mitigate this decrease in performance, where our experiments reflect a minimal performance drop for face anonymization. Our study demonstrates that realistic anonymization can enable privacy-preserving computer vision development with minimal performance degradation across a range of important computer vision benchmarks.