NVIDIA has released NCore v19.6.0, which adds free factory functions for building camera, lidar and external distortion sensor models. NCore previously built those models through from_parameters and maybe_from_parameters static methods, each holding an if/elif table over every concrete subclass. Adding a model meant editing that table, and a model defined outside NCore could not be built at all.
The new functions are camera_model_from_parameters, lidar_model_from_parameters and external_distortion_model_from_parameters, each dispatching on the parameter type it is handed. A register_<kind>_model hook lets models defined outside NCore participate in that dispatch. The static methods remain as deprecated forwarders, so existing calls keep working.
NCore is NVIDIA's data format and library for autonomous vehicle and robotics sensor data, and its first listed feature is neural reconstruction, described on PyPI as designed for data driven neural 3D reconstruction and simulation. 3DGRUT gained NCore v4 dataset support in March, pointing gaussian splatting training directly at NCore sequence JSON files.
ExternalDistortionParameters is new, filling a hierarchy that carried an abstract model base and no abstract parameters base. The three model base classes are now generic in their parameter type and declare get_parameters(), so a CameraModel, LidarModel or ExternalDistortionModel stays usable without narrowing to a concrete model first. CameraModelParameters.transform now returns Self across every override.
Two fixes land, an exact search and compare in closest_index_sorted, and a floored map index in the angles to columns conversion. A team running a custom sensor rig can register its own camera or lidar model and have an NCore pipeline construct it without patching NCore.
NCore v19.6.0 is Apache-2.0 and requires Python 3.8 or newer. It is available now on PyPI as nvidia-ncore.
New to Gaussian Splatting? Start here
