Open Access

Markus RehmannORCiD, Michael BrunnerORCiD, Cristóbal CurioORCiD

• Imbalanced datasets pose a significant challenge in machine learning, where rare situations can significantly impact model performance, particularly in safety-critical applications such as autonomous driving. To ensure models generalize well to new data and real-world scenarios, it is essential to address imbalances during the dataset creation. We propose a novel approach to improve the distribution of human pose datasets by detecting and mitigating rare poses in human motion data. Our method, Out-Of-Distribution Mitigation based on Dynamic Sampling (OODM-DS), employs dimensionality reduction to transform high-dimensional human pose data into a low-dimensional latent space, facilitating efficient analysis of human poses. By analyzing the latent space using density estimation, we propose a dynamic sampling strategy to generate a dataset with a connected simulation. This approach significantly improves the performance of a human pose estimation model on rare poses, while also reducing PCKh metric variance between rare and non-rare poses, making it a promising tool for dataset production in applications such as autonomous driving where safety-critical data can be rare and difficult to obtain.