Anonymizing at-home fitness: enhancing privacy and motivation with virtual reality and try-on

Abstract

Introduction: This study aimed to address privacy concerns associated with video conferencing tools used in home-based exercise training. To that end, a method that could anonymize participants' appearances and exercise environments during at-home fitness sessions was proposed.

Methods: This method combines virtual reality for 3-D human-model rendering using key-points tracking with a virtual try-on system enhanced by UV mapping and instance segmentation. To validate the proposed method, we conducted a user study by recruiting participants to assess effectiveness of virtual reality and virtual try-on in terms of privacy protection, self-confidence, and coaching satisfaction.

Results: Experimental results demonstrated the effectiveness and improved user experience of using virtual reality or virtual try-on in remote fitness, particularly in enhancing privacy protection and self-confidence with statistical significance. However, no significant differences were noted in coaching satisfaction.

Discussion: These findings confirmed the efficacy of our proposed approach. We believe that the proposed approach can significantly contribute to the future of remote fitness training, offering a more secure and engaging environment for users, thereby potentially increasing adherence to fitness regimens and overall physical wellbeing.

Keywords: fitness apps; physical activity; physical health; smart applications; virtual reality exercise; virtual try-on.

Figure 1

Procedure for 3D human modeling with cloth. Given the raw frame (A), the detected 3D key points (B) are mapped to a 3D human model (C) constructed using SMPL model (47). Subsequently, the SNUG model (46) is used to apply clothing to the corresponding 3D human model to render a 3D human model with cloth (D).

Figure 2

Experimental setup with eight installed cameras that are situated at distinct angles and oriented toward the subject.

Figure 3

(A) Segmentation predictor: It aims to generate a semantic segmentation image (I^seg) in real time from a raw image (I^raw) by applying the segmentation map to the UV field predicted by Densepose (48). (B) Overview of the proposed network: The semantic segmentation image (I^seg) is used to try G^top on I^raw, where the garment warping network (GWN) warps G^top to $I_{garment}^{top}$ and inputs it to TON-T. Then, I^raw is divided into the parts for preservation ($I_{preserv}^{top}$) and skin parts for regeneration ($I_{skin}^{top}$), which are used as inputs to TON-T to generate $I_{pred}^{top}$. Similarly, I^seg is employed to try G^bottom on $I_{pred}^{top}$, where the GWN warps G^bottom to $I_{garment}^{bottom}$ and inputs it to TON-T. Subsequently, $I_{pred}^{top}$ is divided into the parts for preservation ($I_{preserv}^{bottom}$) and skin parts for regeneration ($I_{skin}^{bottom}$), which are used as inputs to TON-B to generate $I_{pred}^{top,bottom}$.

Figure 4

Home-based exercise training with VR (3D human-model rendering based on 3D key-points tracking).

Figure 5

Home-based exercise training with UVI-VTON (UV mapping-enhanced instance-segmentation-based virtual try-on network).

Figure 6

Privacy Protection Scale (PPS) consisting of three questions (Question 1, Question 2, Question 3) for all participants. PPS serves as a tool to assess participants' privacy protection levels based on the presence or absence of UVI-VTON and VR. Vertical error bars represent standard deviation for response for each question. *p < 0.05, ***p < 0.001.

Figure 7

Self-confidence and Motivation Scale (SMS) comprising questions 4 and 5 is employed to assess participants' self-confidence and motivation levels with consideration of the presence or absence of UVI-VTON and VR. Additionally, the Coaching Satisfaction Scale (CSS), comprising a single question (Question 6), is used to gauge participants' coaching satisfaction level, taking into account the presence or absence of UVI-VTON and VR. Vertical error bar represents standard deviation for response for each question. *p < 0.05; **p < 0.005.