SynthDa

posted on 01 Dec 2024 by megani

Synthetic action video data generation for CV model training

Brief

This research introduces SynthDa, a user-friendly framework for generating synthetic data in computer vision tasks, particularly action detection.

Unlike traditional methods, SynthDa allows users of varying technical expertise to create and evaluate diverse synthetic datasets, addressing challenges associated with real-world data limitations and the complexities of synthetic data generation. The study demonstrates SynthDa’s utility in evaluating detection performance and provides valuable insights into the impact of synthetic data augmentation on model effectiveness, offering a promising tool for researchers and practitioners.

Members

SIT

NVIDIA

This work is a collaboration between the Centre for Immersification and the SIT-NVIDIA AI Centre (SNAIC)

Outputs

Publications

Megani Rajendran, Chek Tien Tan, Indriyati Atmosukarto, Aik Beng Ng, Joey Lim, Triston Chan Sheen, and Simon See. 2024. Designing a Usable Framework for Diverse Users in Synthetic Human Action Data Generation. In SIGGRAPH Asia 2024 Technical Communications (SA ‘24). Association for Computing Machinery, New York, NY, USA, Article 11. https://doi.org/10.1145/3681758.3697986

Megani Rajendran, Chek Tien Tan, Indriyati Atmosukarto, Aik Beng Ng, Zhihua Zhou, Andrew Grant, and Simon See. 2023. SynthDa: Exploiting Existing Real-World Data for Usable and Accessible Synthetic Data Generation. In SIGGRAPH Asia 2023 Technical Communications (SA ‘23). Association for Computing Machinery, New York, NY, USA, Article 6. https://doi.org/10.1145/3610543.3626168

Megani Rajendran, Chek Tien Tan, Indriyati Atmosukarto, Aik Beng Ng, and Simon See. 2022. SynDa: A Novel Synthetic Data Generation Pipeline for Activity Recognition. In 2022 IEEE International Symposium on Mixed and Augmented Reality Adjunct (ISMAR-Adjunct). IEEE, 373–377. https://doi.org/10.1109/ISMAR-Adjunct57072.2022.00081

Megani Rajendran, Chek Tien Tan, Indriyati Atmosukarto, Aik Beng Ng, Andrew Grant, Eric Cameracci, and Simon See. 2023. Exploring Domain Randomization’s Effect on Synthetic Data for Activity Detection. In 2023 IEEE International Conference on Metaverse Computing, Networking and Applications (MetaCom). IEEE, 139–140. https://doi.org/10.1109/MetaCom57706.2023.00037

Megani Rajendran, Chek Tien Tan, Indriyati Atmosukarto, Aik Beng Ng, and Simon See. 2022. Towards a HOI application for digital home assistants. In 2022 IEEE International Symposium on Mixed and Augmented Reality Adjunct (ISMAR-Adjunct). IEEE, 942–945. https://doi.org/10.1109/ISMAR-Adjunct57072.2022.00211

Megani Rajendran, Chek Tien Tan, Indriyati Atmosukarto, Aik Beng Ng, and Simon See. 2024. Review on synergizing the metaverse and AI-Driven synthetic data: enhancing virtual realms and activity recognition in computer vision. Visual Intelligence 2 (2024). https://link.springer.com/article/10.1007/s44267-024-00059-6

Usable Artifact

[WIP] A web-based + CLI tool for generating synthetic action video data for computer vision model training.