The MVTec MV-ball dataset is a publicly available synthetic dataset for evaluating multi-view 6D pose estimation methods. It features a blue sphere with one red and one green hemisphere on its surfaces, which are separated by 90 degrees. The pose of the sphere can not be determined from views that do not include both views. The dataset only consists of such views such that only holistic multi-view methods can resolve the poses in the dataset reliably.
The dataset consists of 112.042 training images and 3.853 and 3.458 images for evaluation with the easy or the hard split, respectively. For the simple and hard test split, at least 300 and 10 pixels of one hemisphere must be visible, respectively. Each scene contains only one ball to enable a more interpretable evaluation.
More information can be found in our corresponding paper titled “MVTOP: Multi-View Transformer-Based Object Pose-Estimation”.
