Paint a hole on flat pavement and a state-of-the-art depth model will dig it for real. It reads the texture, trusts its priors, and bends geometry that was never there. We call that a 3D mirage, and this is the benchmark built to expose it.

3D-Mirage backs our ECCV 2026 paper, The 3D Mirage: Probing and Taming 3D Hallucinations.

What's inside

468 real-world scenes of painting and street-art illusions: chalk anamorphoses, forced-perspective murals, and large-format billboards that illustrate fake 3D structure on a low curvature surface.
Human-annotated ROI polygons for every illusion, one polygon per surface, with nested cut-outs that mask out real objects and people sitting on the illusion.
Four ROI-centered crops per scene, each sampled so the illusion covers at least 40% of the crop's area, so the illusion stays in view while the surrounding context falls away. We found that context change results in different hallucination behavior. This is the narrow field that a camera on robot or IoT may get, or when getting closer to the illusion.
1,872 full-plus-crop instances, all annotated and checked by hand.

The scenes are real and diverse: indoor and outdoor, on pedestrian walkways, illusions that span more than one surface, and plenty with real objects or people standing on or beside the illusion.

Why it is hard

A model can top the usual depth benchmarks and still fold the instant texture starts lying about geometry. 3D-Mirage puts that failure in the spotlight, then pairs every full image with a tight crop to check whether the model's answer survives once the viewpoint change, context reduce or increase.

How it is scored

Ground-truth depth for a painted illusion is almost impossible to collect at scale, so we drop the requirement. Two reference-free scores run on the annotated flat surface:

DCS (Deviation Composite Score) measures the spurious curvature a model invents inside a region that is physically flat. Lower is calmer.
CCS (Confusion Composite Score) measures how much that same region moves between the full image and its crop. Lower is steadier.

Both read curvature straight off the surface, so a hallucinated bulge that happens to keep depth ordering still gets flagged. These metrics complement existing metrics and allow our method to effectively reduce hallucination (code).

Load it

from huggingface_hub import snapshot_download

path = snapshot_download(repo_id="3dmirage/3D-Mirage", repo_type="dataset")
print("downloaded to:", path)

Every instance carries its full image, the ROI mask(s), and one context-varied crop.

Good for

Stress-testing monocular depth models on illusion robustness.
Measuring how heavily a model leans on scene context.
Training and evaluating hallucination-mitigation methods, such as the Grounded Self-Distillation recipe from our paper.

Citation

@inproceedings{nguyen2026mirage,
  title     = {The 3D Mirage: Probing and Taming 3D Hallucinations},
  author    = {Nguyen, Hoang and Xu, Xiaohao and Huang, Xiaonan},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}

You need to agree to share your contact information to access this dataset

3D-Mirage: Probing 3D Hallucinations in Monocular Depth