Looping Through Color Space: A Simple Augmentation Method to Improve Biased Object Detection

In this work, we address the challenging problem of color-dependent and imbalanced datasets. For many use cases, the training of models based on such data will not generalize well enough and fail even on slight domain variations. This issue is usually addressed by artificially extending the data by manipulating input data or using synthetic data. In this context, we introduce a new augmentation method for extended color mapping from single-channel depth images that reduce color dependency and decrease the amount of annotated data needed for comparable model performance. We found that this method improves the generalization of models for depth-based hand detection on our dataset captured at a manual assembly workspace. Additionally, we validated our results on a publicly available dataset.