Reproduce “SEPose: A Synthetic Event-based Human Pose Estimation Dataset for Pedestrian Monitoring” on OpenResearch

About this paper

Kaustav Chanda, Aayush Atul Verma, Arpitsinh Vaghela, Yezhou Yang, Bharatesh Chakravarthi

Researchers at Arizona State University created SEPose, a large synthetic dataset for event-based human pose estimation, designed for fixed traffic cameras to improve pedestrian monitoring. Experiments demonstrated that models trained on SEPose learned transferable features, indicating its utility for real-world applications despite a sim-to-real performance gap.

Key insights

• Synthetic data generation offers a scalable and cost-effective method to address the significant data scarcity in event-based vision for ITS applications.
• Models trained on SEPose learn meaningful and transferable features for human pose estimation, demonstrating the dataset's potential to generalize to real-world event-based data.
• A notable sim-to-real performance gap exists, suggesting that more accurate simulation of real event camera noise models and exploration of domain adaptation techniques are needed to improve generalization.

Summary by alphaXiv.