MV3DPT is the first benchmark dataset specifically built for the task of "multi-view arbitrary 3D point tracking" released in 2025 by ETH Zurich, Microsoft Mixed Reality and Artificial Intelligence Laboratory, Carnegie Mellon University and other institutions. The relevant research results are "Multi-View 3D Point Tracking", which has been selected for ICCV 2025 (Oral), aims to provide a research basis for "online stable tracking of arbitrary 3D points in dynamic scenes from multiple camera perspectives".

The dataset consists of three subsets:

• MV-Kubric: A synthetic training dataset, which is a multi-view version of the single-view Kubric data, containing about 5,000 synthetic multi-view sequences (multiple camera perspectives).
• Panoptic Studio: A real-world evaluation benchmark covering 10 dynamic scenes such as playing basketball, playing with toys, and juggling, used to test the performance of models in real-world environments.
• DexYCB Multiview: Another real-world evaluation set built on the DexYCB data and incorporating true 3D trajectories of hands and objects.

This dataset covers synthetic and real scenes, multi-view fusion data, can be used for robust prediction under occlusion, is suitable for training and evaluating three-dimensional point tracking models, and is widely used in computer vision and robotics related scenarios.