I'm assuming the poses included in the dataset are those obtained after bundle adjustment optimization. Would it be possible to include the original poses produced by ARKit? This would help with methods that perform camera pose refinement from a set of inaccurate initial poses.