Introduction
-
We present a novel framework for accurate multi-view 2D face localization by leveraging 3D information. We further emphasize the necessity for consistent anonymization across all camera views using our proposed holistic recall.
-
We present a training-free, mesh-based anonymization method yielding complete control during the 3D face replacement step while generating more realistic results than existing state-of-the-art approaches.
-
The images anonymized by our framework can be effectively utilized by downstream methods, as shown through experiments on image quality assessment and downstream face localization.
Related works
Face detection
Image anonymization
Human pose estimation
Methods
Multi-person 3D Mesh Regression
3D Human Representation
Rendering the faces in 2D
Ground truth curation
-
Easy Evaluation Scenario Up to four people in the scene, all wearing surgical masks and hospital scrubs with only a few face obstructions. A total of 1310 faces.
-
Medium Evaluation Scenario Five or six people in the scene with regular face obstructions caused by the position of the surgical lights. A total of 2317 faces.
-
Hard Evaluation Scenario Four people are present in the room. Clinicians additionally wear skull caps and gowns. The surgical lights frequently obstruct the faces in two of the views. A total of 1286 faces.
Experiments
Face localization
Image quality
Results
Face Detection
Method | FID \( \downarrow \) [28] | LPIPS \( \downarrow \) [29] | SSIM \( \uparrow \) [30] |
---|---|---|---|
Blackening | 194.03 | 0.5392 | 0.1864 |
Pixel | 173.34 | 0.4037 | 0.6080 |
Blur | 164.57 | 0.3688 | 0.6014 |
DeepPrivacy [10] | 94.85 | 0.2276 | 0.6294 |
DisguisOR | 35.24 | 0.1341 | 0.8143 |