![]() It is currently the most dominant generative model and the key to its success is the adversarial loss, which forces the generated data distribution to be indistinguishable from the real one. Since then, numerous works have studied ways to stabilize and improve training of GANs to generate high-quality and high-resolution images. Generative Adversarial Networks (GANs) have been proposed to generate realistic images in an unsupervised manner. However, it is specific to the facial attribute setting and is not applicable to more general privacy-sensitive tasks such as action recognition. It achieves impressive results on gray-scale facial pictures with attribute annotations. de-identify faces while preserving facial attributes by fusing faces with similar attributes. Indeed, the low resolution recognition performances of these works were much lower than the state-of-the-arts on high resolution videos, particularly for large-scale video datasets. This means that there is no guarantee that it is optimal for privacy-preserving action recognition. However, although such low resolution downsampling of videos removes scene details, its anonymization strategy is hand-crafted (i.e., it was not learned). All these previous works relied on video downsampling for privacy-protection. further studied the method of learning a better representation space for such very low resolution (e.g., 16x12) videos. developed a two-stream version of, extending it to handle extreme low resolution videos. worked on learning of efficient low resolution video transforms to classify actions from extreme low resolution videos. There can also be hidden backdoors installed by the manufacturer or the government, guaranteeing their access to cameras at one’s home. In the worst case, the users are under the risk of being monitored by a hacker if their cameras or robots at home are cracked. All these create a potential risk of one’s private videos being snatched by someone else. ![]() They sometimes even require network access to high computing power servers, sending potentially privacy-sensitive images/videos. Most computer vision algorithms require loading high resolution images/videos (that contain privacy-sensitive data) to CPU/GPU memory to enable visual recognition. On one hand, we want the camera systems/robots to recognize important events and assist human daily life by understanding its videos, but on the other hand we also want to ensure that they do not intrude the user’s or others’ privacy. Simultaneously, there is an increasing concern in these systems invading the privacy of their users in particular, from unwanted video taking and its sharing. ![]() In this paper, our goal is to create such a system. Ideally, we would like a face anonymizer that can preserve Alex’s privacy (i.e., make his face no longer recognizable as Alex) while at the same time unaltering his actions. However, you do not want your personal assistant to record Alex’s face, because you are concerned about his privacy information since the camera could potentially be hacked. Figure 1: Imagine the following scenario: you would like a personal assistant that can alert you when your adorable child Alex performs undesirable actions, such as eating mom’s make-up or drinking dirty water out of curiosity. For instance, cities are adopting networked camera systems for policing and intelligent resource allocation, individuals are recording their lives using wearable devices, and service robots at homes and public places are becoming increasingly available and popular. See the project page for a demo video and more results.Ĭomputer vision technology is enabling automated understanding of large-scale visual data, making it crucial for many societal applications with ubiquitous cameras. We experimentally confirm the benefit of our approach compared to conventional hand-crafted video/face anonymization methods including masking, blurring, and noise adding. The end result is a video anonymizer that performs a pixel-level modification to anonymize each person’s face, with minimal effect on action detection performance. We use an adversarial training setting in which two competing systems fight: (1) a video anonymizer that modifies the original video to remove privacy-sensitive information (i.e., human face) while still trying to maximize spatial action detection performance, and (2) a discriminator that tries to extract privacy-sensitive information from such anonymized videos. In this paper, we propose a new principled approach for learning a video face anonymizer. On one hand, we want the camera systems/robots to recognize important events and assist human daily life by understanding its videos, but on the other hand we also want to ensure that they do not intrude people’s privacy. There is an increasing concern in computer vision devices invading the privacy of their users by recording unwanted videos.
0 Comments
Leave a Reply. |