This dataset is a subset of 74 videos from the multimodal in-the-wild dataset MuSe-CAR. It contains 1 124 video frames showing human-vehicle interactions across all MuSe topics and 6 146 labels (bounding boxes). The pre-defined training, development and test partitions are also provided. The purpose of this dataset is to support research in the field of automatic recognition and detection of automotive parts in a natural context. It provides labels for 29 interior and exterior vehicle regions during human-vehicle interaction. It also enables benchmarking and cross-corpus transfer learning, as demonstrated in GoCarD (A Generic, Optical Car Part Recognition and Detection). The footage captures many "in-the-wild" characteristics, including a range of shot sizes, camera motion, moving objects, a wide variety of backgrounds and different interactions. The MuSe data set can only be used for research purposes (see below).