MisCue3D is a diagnostic dataset for studying context dependence in visually grounded 3D pose estimation. The dataset is designed to test whether a model's predicted pose, depth, or scale changes when the target object itself is held fixed but the surrounding visual context is changed.
This repository contains rendered observations and metadata from controlled Unreal Engine experiments. It does not contain Unreal Engine project files, .uasset files, source meshes, source… See the full description on the dataset page:
https://huggingface.co/datasets/NeurIPSMisCue3D/2026nipsdata.
