Pedestrian Attribute Editing for Gait Recognition and Anonymization (2303.05076v2)

Abstract: As a kind of biometrics, the gait information of pedestrians has attracted widespread attention from both industry and academia since it can be acquired from long distances without the cooperation of targets. In recent literature, this line of research has brought exciting chances along with alarming challenges: On the positive side, gait recognition used for security applications such as suspect retrieval and safety checks is becoming more and more promising. On the negative side, the misuse of gait information may lead to privacy concerns, as lawbreakers can track subjects of interest using gait characteristics even under face-masked and clothes-changed scenarios. To handle this double-edged sword, we propose a gait attribute editing framework termed GaitEditor. It can perform various degrees of attribute edits on real gait sequences while maintaining the visual authenticity, respectively used for gait data augmentation and de-identification, thereby adaptively enhancing or degrading gait recognition performance according to users' intentions. Experimentally, we conduct a comprehensive evaluation under both gait recognition and anonymization protocols on three widely used gait benchmarks. Numerous results illustrate that the adaptable utilization of GaitEditor efficiently improves gait recognition performance and generates vivid visualizations with de-identification to protect human privacy. To the best of our knowledge, GaitEditor is the first framework capable of editing multiple gait attributes while simultaneously benefiting gait recognition and gait anonymization. The source code of GaitEditor will be available at https://github.com/ShiqiYu/OpenGait.

• The paper introduces GaitEditor, a unified framework that uses attribute-conditioned edits via GANs to enhance gait recognition and achieve effective anonymization.
• It employs a two-stage training process with StyleGAN-ADA and an Attribute-Identity Encoder to manipulate multiple gait attributes with fine-grained control.
• Experimental results show a +1.60% gain in rank-1 accuracy for recognition and a -57.34% drop in accuracy for anonymization on the CCPG dataset.

Overview of "Pedestrian Attribute Editing for Gait Recognition and Anonymization"

This paper introduces "GaitEditor," a novel framework designed to address dual purposes in pedestrian gait analysis: improving gait recognition and ensuring anonymization to protect privacy. GaitEditor is an innovative solution within the field of gait biometrics, which enables attribute-conditioned edits on real gait sequences, harnessing the capabilities of generative adversarial networks (GANs).

Key Contributions

1. Unified Framework for Gait Editing: GaitEditor is unique in its ability to edit multiple gait attributes simultaneously, improving visual authenticity while serving both recognition and anonymization objectives. It efficiently uses a style-based GAN for creating a semantically rich latent space, allowing fine-grained manipulation of pedestrian attributes.
2. Enhanced Gait Recognition: Through data augmentation via attribute manipulation, GaitEditor enhances the performance of several gait recognition models. It achieves notable improvements in rank-1 accuracy across various models and datasets, such as an increase of +1.60% on the CCPG dataset for the baseline GaitBase model.
3. Effective Gait Anonymization: For anonymization, GaitEditor degrades identity consistency while maintaining natural appearances, thus protecting individual privacy without compromising the overall silhouette texture. It significantly disrupts recognition accuracy (e.g., -57.34% on the CCPG dataset), illustrating its effectiveness in anonymizing personal identity.

Methodology

The framework employs a two-stage training process. Initially, the StyleGAN-ADA model constructs a latent space that captures diverse gait attributes from silhouette images. Subsequently, an Attribute-Identity Encoder projects real gait sequences into this latent space. This encoder is designed to maintain temporal consistency and identity features while enabling viewpoint translation.

Gait Attributes Manipulation:

GaitEditor identifies and exploits semantic directions in the latent space that correspond to various pedestrian attributes such as clothing style, body size, and perceived age. Through controlled manipulation of these directions, GaitEditor can finely tune these attributes according to predefined parameters.

Experimental Validations

• User Study: A user paper confirmed the perceptual authenticity of edited attributes, demonstrating significant accuracy in recognizing edited features, especially among participants with gait analysis experience.
• Extensive Benchmarks: GaitEditor was rigorously tested on prominent datasets like OU-MVLP, CCPG, and Gait3D. Results indicate improvements in both recognition accuracy and anonymization metrics.

Implications and Future Directions

GaitEditor's dual capability presents important practical applications in surveillance and privacy preservation. Its adaptability to various datasets and recognition models showcases its robust potential.

Future research can expand on gait anonymization, exploring reversible transformations for privacy-conserving data sharing. Additionally, further exploration into unsupervised learning paradigms may enhance scalability, addressing challenges in the collection and annotation of vast gait datasets.

In summary, GaitEditor emerges as a significant contribution to pedestrian gait analysis, offering a balanced approach to enhancing recognition capabilities while safeguarding privacy.