3D occupancy prediction — estimating which voxels in a scene are occupied and by what — underpins autonomous driving perception. But existing methods require precise camera calibration and in-domain training data. Deploy in a new city with different cameras, and performance collapses.
Note (2026-11-15; read against the paper's full text): the essay says the calibration dependence of earlier methods was a training artifact and that the scene supplies the geometric constraints without the camera. The paper builds its model on pretrained visual geometry foundation models that themselves predict camera quantities (for two of the compared variants it uses their predicted camera intrinsics together with a metric depth estimator to fix scale), and it reports that the model outperforms the visual-geometry baselines and is competitive with in-domain self-supervised methods while being zero-shot. It states that the gap to fully supervised in-domain performance remains substantial and that the results remain behind more recent self-supervised methods. The reading about calibration being a training artifact is the essay's.
The authors (arXiv:2603.23502) introduce OccAny, which predicts metric occupancy with segmentation features from uncalibrated, out-of-domain scenes. It works with sequential, monocular, or surround-view inputs without retraining.
Two enabling techniques: Segmentation Forcing, which conditions occupancy prediction on instance-level segmentation and improves quality while enabling mask-level outputs. And Novel View Rendering, which synthesizes additional viewpoints at test time to fill geometric gaps — particularly effective for monocular inputs where only one perspective is available.
The through-claim: the calibration dependence of prior 3D prediction methods was a training artifact, not a fundamental requirement. The geometric information needed for occupancy prediction exists in the image content, not in the camera parameters. OccAny demonstrates that with the right architecture, the scene itself provides sufficient geometric constraints, even without knowing how the camera was pointed.