Recoverable Anonymization for Pose Estimation: A Privacy-Enhancing Approach

TL;DR

This work introduces a reversible privacy-enhancing system for human pose estimation that simultaneously obfuscates SPI and preserves pose-relevant features. It deploys three jointly trained components—a privacy-enhancing generator $G_X$, a privacy-recovery generator $G_Y$, and a pose estimator $P$—with losses $L_{enhance}$, $L_{recovery}$, and $L_{PE}$ to ensure privacy, recoverability by authorized users, and strong HPE performance. The framework preserves contextual information and enables edge deployment by operating near cameras, while experiments on MPII and COCO demonstrate robust privacy improvement, effective SPI recovery, and competitive pose accuracy, including about a 3–10% gain in key metrics when jointly trained. This approach advances privacy in vision systems by integrating reversibility, context preservation, and lightweight deployment, with practical implications for surveillance and law enforcement applications.

Abstract

Human pose estimation (HPE) is crucial for various applications. However, deploying HPE algorithms in surveillance contexts raises significant privacy concerns due to the potential leakage of sensitive personal information (SPI) such as facial features, and ethnicity. Existing privacy-enhancing methods often compromise either privacy or performance, or they require costly additional modalities. We propose a novel privacy-enhancing system that generates privacy-enhanced portraits while maintaining high HPE performance. Our key innovations include the reversible recovery of SPI for authorized personnel and the preservation of contextual information. By jointly optimizing a privacy-enhancing module, a privacy recovery module, and a pose estimator, our system ensures robust privacy protection, efficient SPI recovery, and high-performance HPE. Experimental results demonstrate the system's robust performance in privacy enhancement, SPI recovery, and HPE.

Figures (5)

• Figure 1: Motivation for our privacy-enhancing system. (a). Conventional surveillance systems are susceptible to leaks of SPI, which can be exploited for illicit surveillance and criminal activities. (b). Our system not only safeguards SPI against information misuse but also supports HPE. The privacy-enhanced images retain functionality for routine monitoring, while SPI remains recoverable by authorized personnel.
• Figure 2: The complete pipeline of our proposed system. It contains a privacy-enhancing module $G_\mathcal{X}$ erasing private information, a module $G_\mathcal{Y}$ recovering the removed private information, two discriminators $D_\mathcal{X}, D_\mathcal{Y}$ for distinguishing the generated portraits, and a pose estimator $\mathcal{P}$ implementing pose estimation. denotes the trainable modules, and denotes the frozen modules.
• Figure 3: Qualitative comparison on privacy-enhanced portraits. (a) original portraits; (b)/(d) conventional desensitized portraits via blurring/pixelation; (c)/(e) privacy-enhanced portraits guided by blurring/pixelation. Enlarge for details.
• Figure 4: Qualitative results of the privacy-recovered portraits. (a) original portraits; (b)/(c) the portraits recovered from the privacy-enhanced portraits guided by blurring/pixelation. Enlarge for details.
• Figure 5: Qualitative results of the privacy-enhanced portraits guided by Gaussian noise. (a) conventional desensitized portraits; (b) privacy-enhanced portraits guided by Gaussian noise addition.