SMPLy Private: From Masks to Meshes in Action Recognition

Nidhish Shah, Asfandyar Azhar, Shaurjya Mandal, Yongjie Jessica Zhang

Keywords: Action Recognition, Computer Vision, Body Mesh Recovery, Dataset Augmentation, Video Data Processing

Abstract: In this paper, we introduce Mask2Mesh (M2M), a novel privacy-preserving data augmentation framework that effectively bridges the realism gap seen in synthetic-based action recognition methods. Traditional privacy-enhancing techniques, such as feature masking and synthetic data supplementation, tend to degrade data quality and reduce model performance. In contrast, our method leverages the SMPL-X model to replace real humans with detailed 3D meshes in video data, preserving the subtle nuances of human movement and expressions that are crucial for accurate action recognition. By augmenting real data with superimposed meshes, M2M simplifies both pre-training and fine-tuning processes, without introducing the overheads and biases typically associated with synthetic data. Empirical results show that our approach achieves performance within 0.5\% of models trained on unmodified video data, proving that overlaying meshes leads to no significant performance loss in action recognition tasks. This work presents a practical solution for data anonymization without compromising accuracy, offering valuable insights for more efficient and scalable video data processing techniques in computer vision and action recognition.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 8122