Object Neural Field Registration (ONR) Dataset

The Object Neural Field Registration (ONR) Dataset is a dataset of raw data and pre-trained 3D neural field models for conducting neural field registration experiments. This is a simulated data benchmark for calculating the pose of objects within a scene using only neural field model representations of both. Data consists of simulated RGB+Depth images, instance segmentation images, object instance poses, camera poses, and trained neural field models designed to work within the sdfstudio framework.\nLineage: Data was created through the Omniverse Isaac Sim simulator to create object and scene level data.\nData for object models were created using pre-defined camera poses that form a circle around the object model of interest which has been placed within an empty void.\nData for scene models were created by manually piloting camera within simulated scenes and recording camera poses as part of capture.\nSpecialised data writing scripts were created to save data from Isaac Sim in a data format compatible with sdfstudio neural field training software.\nTrained models shared as part of this dataset were trained using sdfstudio code on CSIRO HPC machines.