PIVOT Dataset

PIVOT (Pose, Intrinsics and Viewpoint Oriented Testbed) is a multi-trajectory drone dataset designed for evaluating NeRF and 3D Gaussian Splatting under capture conditions closer to real-world robotics and autonomous systems. It provides multiple structured camera trajectories per scene together with measured sensor-based poses and COLMAP-optimized poses.

PIVOT Toolchain

The dataset is accompanied by the PIVOT GitHub repository, which provides the processing, visualization, dataset-export, and benchmarking tools, the dataset is also integrated with Nerfstudio on this branch Nerfstudio PIVOT integration GitHub repository.

Using PIVOT, the released scenes can be:

• Interactively visualized, including the sparse point cloud, measured and COLMAP-optimized camera trajectories, pose errors, trajectory statistics, and directed pose Chamfer distances.
• Exported directly to Nerfstudio format for training and evaluation with Nerfstudio-compatible reconstruction methods.

See the PIVOT repository README for installation instructions, visualization commands, Nerfstudio export, and benchmark execution.

Benchmarks & Paper

PIVOT defines benchmarks for novel-view generalization, measured vs. optimized camera poses, and calibrated vs. optimized camera intrinsics.

Benchmark methodology, quantitative results, qualitative results, and analysis are provided in the PIVOT paper.

Known Issues

• Church scene — rocket_upward: the rocket-upward trajectory was not successfully registered by COLMAP. The trajectory and its measured poses remain part of the dataset, but COLMAP-optimized poses are unavailable for this trajectory. This is planned to be addressed in an upcoming dataset version.

Acknowledgements

Special thanks to Clare County Council and Bunratty Castle & Folk Park, including everyone who assisted with and facilitated the drone captures at the site. Their support and cooperation helped make this research possible.

License

The PIVOT dataset is released under the CC BY-NC 4.0 license.