README.md

---
pretty_name: "MD-3K: MultiDepth-3K Benchmark"
language:
- en
license: other
size_categories:
- 1K<n<10K
task_categories:
- depth-estimation
tags:
- depth-estimation
- monocular-depth-estimation
- ordinal-depth
- spatial-reasoning
- transparent-scenes
- glass
- multi-layer-depth
- geometric-ambiguity
- eccv-2026
---

# MD-3K: MultiDepth-3K Benchmark

MD-3K is a real-world diagnostic benchmark for probing **geometric ambiguity** in monocular depth estimation. It focuses on transparent scenes where a single camera ray can contain two visually present and geometrically valid surfaces: a transparent foreground surface and the visible background behind it. Since a standard monocular depth model returns one scalar depth per pixel, MD-3K evaluates which valid depth layer a model reports and whether paired hypotheses can jointly satisfy both layer-specific ordinal relations.

This dataset accompanies the ECCV 2026 paper:

> **One Scene, Two Depths: Probing Geometric Ambiguity in Monocular Foundation Models**  
> Xiaohao Xu, Feng Xue, Xiang Li, Haowei Li, Shusheng Yang, Tianyi Zhang, Matthew Johnson-Roberson, Xiaonan Huang.

Code repository: `https://github.com/Xiaohao-Xu/Ambiguity-in-Space`

## Release files

The canonical archive in this release is:

```text
MD-3K.zip
```

The expected SHA256 checksum is:

```text
3e9627d3aca6886a4449fd149b7ec791d13943d423fc5ee14f611c986bdecd29  MD-3K.zip
```

After extraction, the dataset should follow this layout:

```text
MD-3K/
├── images/              # RGB images
├── masks/               # ambiguous/transparent-region masks
└── annotations.json     # sparse point-pair ordinal annotations
```

## Dataset summary

MD-3K contains:

- 3,161 RGB images curated for transparent-scene ambiguity analysis.
- Ambiguous-region masks.
- Sparse point-pair annotations in ambiguous regions.
- Two layer-specific ordinal targets:
  - **Layer 1:** transparent foreground surface.
  - **Layer 2:** visible background behind the transparent surface.
- Same/Reverse subset structure:
  - **Same:** the two valid layers induce the same ordinal relation.
  - **Reverse:** the two valid layers induce conflicting ordinal relations.

The **Reverse** subset is the key diagnostic setting: a duplicated single-depth hypothesis cannot satisfy both valid layer relations.

## Download

Download the archive with `huggingface_hub`:

```python
from huggingface_hub import hf_hub_download

zip_path = hf_hub_download(
    repo_id="xiaohaox/MultiDepth-3K-Dataset",  # change this if the dataset is released under another HF namespace
    filename="MD-3K.zip",
    repo_type="dataset",
)
print(zip_path)
```

Extract locally:

```python
from pathlib import Path
from zipfile import ZipFile

zip_path = Path(zip_path)
out_dir = Path("data")
out_dir.mkdir(parents=True, exist_ok=True)

with ZipFile(zip_path, "r") as zf:
    zf.extractall(out_dir)

print(f"Extracted to: {out_dir.resolve()}")
```

Expected local path for the GitHub evaluation scripts:

```text
data/MD-3K/
```

## Integrity check

After downloading, verify the archive.

Linux:

```bash
sha256sum -c checksums.txt
```

macOS:

```bash
shasum -a 256 MD-3K.zip
```

Windows PowerShell:

```powershell
Get-FileHash -Algorithm SHA256 MD-3K.zip
```

The expected SHA256 hash is:

```text
3E9627D3ACA6886A4449FD149B7EC791D13943D423FC5EE14F611C986BDECD29
```

## Annotation schema

The canonical annotation file is `annotations.json`. It maps an image path to a list of sparse point-pair annotations. A typical record is:

```json
{
  "images/example.jpg": [
    {
      "point1": [x1, y1],
      "point2": [x2, y2],
      "foreground_label": "near_or_far",
      "background_label": "near_or_far",
      "subset": "same_or_reverse"
    }
  ]
}
```

If a compact legacy schema is used, it may contain a single `label` field. In that case, follow the repository evaluation protocol and document the mapping explicitly. Do **not** describe `label = 1` as SRA(1) or `label = 2` as SRA(2): SRA(1) and SRA(2) are evaluation targets for the foreground and background layers, not merely dataset label values.

Coordinate convention: point coordinates are stored as `[x, y]`. When indexing an image or depth array with NumPy/OpenCV, use `[y, x]`.

## Depth convention

Before comparing a predicted depth map to MD-3K ordinal labels, verify the saved prediction convention.

Some models save depth-like outputs, where larger values indicate farther points. Other models save inverse-depth or disparity-like outputs, where larger values indicate nearer points. MD-3K evaluation requires a consistent near/far convention. If a prediction is inverse depth or disparity, convert it to depth, or equivalently flip the ordinal comparison before computing SRA(1), SRA(2), depth-layer preference, or ML-SRA.

## Metrics

### SRA(1)

Spatial Relationship Accuracy with respect to the transparent foreground layer.

### SRA(2)

Spatial Relationship Accuracy with respect to the visible background layer.

### Depth-layer preference

```text
alpha = SRA(2) - SRA(1)
```

Positive `alpha` indicates stronger background-layer preference. Negative `alpha` indicates stronger foreground-layer preference.

### ML-SRA

Multi-Layer Spatial Relationship Accuracy evaluates whether a paired output, such as RGB and LVP predictions, jointly satisfies both layer-specific ordinal relations.

In the paper protocol, RGB and LVP are assigned according to the model's **RGB depth-layer preference at the benchmark level**:

- If RGB prefers layer 1, assign RGB to layer 1 and LVP to layer 2.
- If RGB prefers layer 2, assign RGB to layer 2 and LVP to layer 1.

This is not a per-image oracle and should not be described as automatic test-time layer selection. Alternative pairing strategies may be useful future improvements, but they should be reported separately.

## Associated code and LVP

The evaluation and analysis code is maintained in the GitHub repository:

```text
https://github.com/Xiaohao-Xu/Ambiguity-in-Space
```

Laplacian Visual Prompting (LVP) is a deterministic, training-free input transformation. It has no learned parameters and no checkpoint, so it is not released here as a Hugging Face model. The intended use is to apply the transform to the input image, map the result back to a standard RGB image representation, and pass it through the same depth-model processor used for the original RGB input.

## Recommended use

Use MD-3K to:

- characterize depth-layer preference in monocular depth models,
- compare RGB and LVP behavior,
- evaluate paired-hypothesis complementarity with ML-SRA,
- analyze Same and Reverse subsets separately.

Do not use MD-3K to claim dense metric multi-layer depth recovery unless additional dense ground truth and evaluation are provided.

## Limitations

MD-3K is a focused diagnostic benchmark. It is not intended to exhaustively cover all transparency, reflection, refraction, material, or open-world 3D ambiguity cases. Its annotations are sparse ordinal point-pair labels rather than dense metric multi-layer depth maps. It should be used for controlled analysis of model behavior under layered transparent-scene ambiguity, not as a standalone guarantee of robotic safety or deployable open-world depth robustness.

## License and data terms

This dataset card uses `license: other` because the release contains dataset annotations/masks and RGB imagery curated from upstream transparent/glass-scene data. Please follow the terms of this MD-3K release and the applicable upstream data terms. If you decide to publish the complete archive under a specific standard license, update both the YAML metadata and this section consistently.

## Citation

If you use MD-3K, please cite:

```bibtex
@inproceedings{xu2026onescenetwodepths,
  title={One Scene, Two Depths: Probing Geometric Ambiguity in Monocular Foundation Models},
  author={Xu, Xiaohao and Xue, Feng and Li, Xiang and Li, Haowei and Yang, Shusheng and Zhang, Tianyi and Johnson-Roberson, Matthew and Huang, Xiaonan},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2026}
}
```

Also, the segmentation labels are aourced from GDD dataset. So please cite:
```bibtex
@inproceedings{Mei_2020_CVPR,
    author = {Mei, Haiyang and Yang, Xin and Wang, Yang and Liu, Yuanyuan and He, Shengfeng and Zhang, Qiang and Wei, Xiaopeng and Lau, Rynson W.H.},
    title = {Don't Hit Me! Glass Detection in Real-World Scenes},
    booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month = {June},
    year = {2020}
}
```