[Paper Review] Does Image Anonymization Impact Computer Vision Training?

Håkon Hukkelås, Frank Lindseth|arXiv (Cornell University)|Jun 8, 2023

Privacy-Preserving Technologies in Data4 citations

TL;DR

This paper investigates the impact of traditional and realistic image anonymization on computer vision model training, using face and full-body anonymization on BDD100K, Cityscapes, and COCO datasets. It finds that realistic anonymization preserves model performance significantly better than traditional blurring, especially for face-related tasks, while full-body anonymization still causes notable performance drops even with realistic methods.

ABSTRACT

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.

Motivation & Objective

• To assess how traditional and realistic image anonymization affect performance in key computer vision tasks like detection, instance segmentation, and pose estimation.
• To investigate whether realistic anonymization—using generative models to synthesize identities—can preserve data utility better than traditional obfuscation (e.g., blurring, masking).
• To evaluate the impact of full-body anonymization, which is often required for privacy but underexplored in model training contexts.
• To determine whether realistic anonymization can enable privacy-preserving computer vision development with minimal performance degradation.

Proposed method

• Anonymized public computer vision datasets (BDD100K, Cityscapes, COCO) using both traditional methods (blurring, masking) and realistic anonymization via DeepPrivacy2’s generative models.
• Trained object detection, instance segmentation, and keypoint detection models on anonymized training data using ResNet and R-CNN architectures.
• Evaluated model performance on original, unanonymized validation sets to measure recognition accuracy drops.
• Used automatic detectors (DSFD for faces, ViTPose for keypoints) to identify and anonymize individuals, though this introduced potential annotation ambiguity.
• Applied histogram equalization to reduce visual mismatches between synthesized identities and global image context.
• Focused on anonymization of faces and full bodies, with particular attention to context consistency and resolution constraints.

Figure 3 : DeepPrivacy2 [ 23 ] anonymizes one instance at a time, then paste each synthesized individual into the original image. For our experiments, detection is not performed, as segmentation masks are defined from pre-defined annotations (see Sec. 3.1 ). Note that the generator relies on keypoin

Experimental results

Research questions

• RQ1 Does realistic anonymization preserve computer vision model performance better than traditional anonymization methods like blurring or masking?
• RQ2 How does full-body anonymization impact model performance across detection, segmentation, and pose estimation tasks?
• RQ3 To what extent does face anonymization degrade performance in models trained on datasets with varying face prevalence?
• RQ4 Can realistic anonymization mitigate performance degradation while maintaining privacy in autonomous vehicle datasets?

Key findings

• Traditional face anonymization via blurring causes a significant performance drop in keypoint detection models on the COCO dataset, where faces are more prevalent.
• Realistic face anonymization reduces the performance drop in keypoint detection to minimal levels, demonstrating its superiority over traditional methods.
• Full-body anonymization via traditional blurring severely impairs model performance across all evaluated datasets, particularly in detection and segmentation tasks.
• Realistic full-body anonymization notably alleviates performance degradation compared to traditional methods, though not to the same extent as for face anonymization.
• The performance drop for detecting 'rider' and 'bicycle' classes is likely due to detection overlaps, not anonymization per se, as other vehicle classes remain unaffected.
• Current generative models for anonymization struggle with context alignment and resolution mismatches, producing unnatural results when bodies are small or poorly localized.

Original