SynDa: A Novel Synthetic Data Generation Pipeline for Activity Recognition

Megani Rajendran, Chek Tien Tan, Indriyati Atmosukarto, Aik Beng Ng, Simon See

Abstract: The task of classifying or predicting the activity performed by hu-mans is called human activity recognition. Many existing models aim to solve the problem of activity recognition in this field, but we recognise the lack of real data can have a great impact on the effectiveness of such models. In this paper, we introduce SynDa, a first-of-its-kind streamlined semi-automated pipeline for synthetic data generation built using photorealistic rendering and AI pose estimation, that harvests existing real-life video datasets to create new large-scale datasets. The synthetic data can augment real data to train models robustly and overcome the lack of and associated costs to acquire more real data. Preliminary work indicates that combining real data and synthetic video data generated using this pipeline to train models presents a mAP of 32.35%, while a model trained on real data presented 29.95%.