Realistic Full-Body Anonymization with Surface-Guided GANs

Hukkelås, Håkon; Smebye, Morten; Mester, Rudolf; Lindseth, Frank

doi:10.1109/wacv56688.2023.00148

Cited by 9 publications

(5 citation statements)

References 48 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Dvořáček et al [21] and Zhou et al [22] demonstrated that the performance, particularly object detection and instance segmentation, are adversely impacted when the urban scenes dataset is anonymized. Hukkelås et al [23] reported declining instance segmentation performance on the anonymized autonomous driving dataset. Therefore, recent research efforts have been focused on methods aimed at substituting identifiable objects with synthetic images, effectively impeding the potential for data restoration while preserving its intrinsic value [24].…”

Section: ) Traditional Methodsmentioning

confidence: 99%

“…The CSE-detector neural network can estimate the correspondence between a body shape and the model of the human body surface within the image. Based on this estimated body surface model and their posture, a GAN can be employed to generate virtual individuals [23]. Another approach involves altering only the target features.…”

Section: ) Traditional Methodsmentioning

confidence: 99%

“…In the case of humans, not only face information but also data like gait and silhouettes can be utilized for identity recognition. Several studies investigate the impact of full-body anonymization to address privacy concerns [23], [24]. Similarly, with vehicles, apart from license plates, aspects such as vehicle color, type, brand, or accessories can facilitate personal identification.…”

Section: Privacy-preserving In Large Scale Visual Datasetsmentioning

confidence: 99%

See 2 more Smart Citations

Balancing Privacy and Accuracy: Exploring the Impact of Data Anonymization on Deep Learning Models in Computer Vision

Lee,

You

2024

IEEE Access

View full text Add to dashboard Cite

Computer vision has become indispensable in various applications, including autonomous driving, medical imaging, security and surveillance, robotics, and pattern recognition. In recent years, the quality of training data has emerged as a critical factor for ensuring effectiveness in real-world scenarios. However, the increasing stringency of privacy regulations in various regions necessitates careful handling of collected images for computer vision. Personal information within images is typically anonymized by applying anonymization patterns to remove it. Empirical findings underscore the significant influence of data quality on the training of deep learning models. Striking the right balance between privacy and recognition performance becomes paramount. Therefore, it is essential to understand how the anonymization of image datasets affects deep learning model performance. In this paper, we thoroughly analyze the effects of different anonymization techniques on the performance of deep learning-based models in computer vision tasks, with a particular emphasis on presenting a model-centric perspective, such as the type of deep learning model and the number of parameters. We aim to provide valuable insights and guidelines for selecting the optimal level of anonymization that strikes a balance between recognition accuracy and privacy protection.

show abstract

Section: ) Traditional Methodsmentioning

confidence: 99%

Section: ) Traditional Methodsmentioning

confidence: 99%

Section: Privacy-preserving In Large Scale Visual Datasetsmentioning

confidence: 99%

See 1 more Smart Citation

Balancing Privacy and Accuracy: Exploring the Impact of Data Anonymization on Deep Learning Models in Computer Vision

Lee,

You

2024

IEEE Access

View full text Add to dashboard Cite

show abstract

“…preserving facial expression) but offer no formal guarantee of removing the original identity from the image, making them vulnerable to adversarial attacks. A few methods explore anonymizing the full-body [4,23,26,38], where the current state-of-the-art [23,24] can generate convincing full-bodies given sparse keypoints [24] or dense pose annotations [23]. Finally, some methods insert adversarial perturbation in the image, which is invisible to the human eye but able to fool face recognition systems [46].…”

Section: Related Workmentioning

confidence: 99%

“…gait, clothes, ear, body shape), often requiring fullbody anonymization to protect privacy. A few studies explore the impact of full-body anonymization [23,26], where they find it to improve over traditional methods. However, they rely on automatic detection methods, which opens the question if the performance degradation is due to detection errors or the anonymization model.…”

Section: Introductionmentioning

confidence: 99%

Image Inpainting with Learnable Feature Imputation

Hukkelås

Lindseth

Mester

2021

Lecture Notes in Computer Science

View full text Add to dashboard Cite

A regular convolution layer applying a filter in the same way over known and unknown areas causes visual artifacts in the inpainted image. Several studies address this issue with feature re-normalization on the output of the convolution. However, these models use a significant amount of learnable parameters for feature re-normalization [ 41 , 48 ], or assume a binary representation of the certainty of an output [ 11 , 26 ]. We propose (layer-wise) feature imputation of the missing input values to a convolution. In contrast to learned feature re-normalization [ 41 , 48 ], our method is efficient and introduces a minimal number of parameters. Furthermore, we propose a revised gradient penalty for image inpainting, and a novel GAN architecture trained exclusively on adversarial loss. Our quantitative evaluation on the FDF dataset reflects that our revised gradient penalty and alternative convolution improves generated image quality significantly. We present comparisons on CelebA-HQ and Places2 to current state-of-the-art to validate our model. (Code is available at: github.com/hukkelas/DeepPrivacy . Supplementary material can be downloaded from: folk.ntnu.no/haakohu/GCPR_supplementary.pdf ) Electronic supplementary material The online version of this chapter (10.1007/978-3-030-71278-5_28) contains supplementary material, which is available to authorized users.

show abstract