Exploiting Self-Imposed Constraints on RGB and LiDAR for Unsupervised Training

Hand detection on single images is an intensively researched area, and reasonable solutions are already available today. However, fine-tuning detectors within a specific domain remains a tedious task. Unsupervised training procedures can reduce the effort required to create domain-specific datasets and models. In addition, different modalities of the same physical space, here color and depth data, represent objects differently and thus allow for exploitation. We introduce and evaluate a training pipeline to exploit the modalities in an unsupervised manner. The supervision is omitted by choosing suitable self-imposed constraints for the data source. We compare our training results with ground truth training results and show that with these modalities, the domain can be extended without a single annotation, e.g., for detecting colored gloves.