---
license: cc-by-nc-sa-4.0
task_categories:
- depth-estimation
tags:
- depth
- monocular-depth
- affine-invariant
pretty_name: DinoDepth
size_categories:
- 100K<n<1M
configs:
- config_name: train
  default: true
  data_files:
  - split: train
    path: train/*.parquet
- config_name: nyuv2
  data_files:
  - split: test
    path: eval/nyuv2-*.parquet
- config_name: kitti
  data_files:
  - split: test
    path: eval/kitti-*.parquet
---

![DinoDepth](cover.png)

# DinoDepth

A large, harmonized, pre-shuffled corpus of **358,905 (image, depth) pairs** for training
monocular **affine-invariant** depth models. Five complementary sources — indoor, driving, robotics,
aerial, and multi-view-stereo — are decoded to a **single planar-depth convention**, packed into
uniform ~1 GB Parquet shards, and **globally shuffled** so any shard is a representative sample of
the whole.

## Schema

| column | type | description |
|---|---|---|
| `dataset` | string | source dataset tag |
| `height`, `width` | int32 | native resolution of the stored arrays |
| `image` | binary | JPEG bytes (RGB) |
| `depth` | binary | 16-bit PNG &mdash; **`depth = depth_png * depth_scale`**, `0` = invalid |
| `depth_scale` | float32 | metres per PNG level (arbitrary per-scene units for multi-view stereo) |

```python
import io, numpy as np
from PIL import Image
from datasets import load_dataset

ds = load_dataset("blanchon/dinodepth-dataset", split="train", streaming=True)
row = next(iter(ds))
rgb   = Image.open(io.BytesIO(row["image"]))                                  # RGB
depth = np.asarray(Image.open(io.BytesIO(row["depth"]))) * row["depth_scale"]  # m; 0 = invalid
```

## Composition

| source | samples | domain | depth |
|---|---:|---|---|
| TartanAir | 186,693 | robotics / aerial (synthetic) | metric |
| BlendedMVS | 74,838 | multi-view stereo (real images) | non-metric |
| IRS | 57,819 | indoor (synthetic) | metric |
| Hypersim | 26,912 | indoor (synthetic) | metric |
| Virtual KITTI 2 | 12,643 | driving (synthetic) | metric |
| **total** | **358,905** | | |

## Harmonization

Every source is decoded to **planar depth `Z`**, with invalid pixels (sky, far-plane saturation,
unreconstructed, non-finite) set to `0`:

- **Hypersim** &mdash; ray distance &rarr; planar `Z` (per-pixel, focal 886.81 px).
- **Virtual KITTI 2** &mdash; native depth / 100 (cm &rarr; m).
- **IRS** &mdash; disparity &rarr; depth (`48 / disparity`).
- **TartanAir** &mdash; native metres; sky masked.
- **BlendedMVS** &mdash; per-scene multi-view-stereo depth (arbitrary, non-metric scale).

Depth is stored at native resolution as 16-bit PNG with a per-image `depth_scale`. Metric sources
are clamped to a 100 m far plane; the multi-view-stereo source keeps its per-scene scale &mdash; a
scale-and-shift-invariant objective absorbs it, so the sources mix directly.

## Splits

- **`train`** &mdash; the full 358,905-image corpus, globally shuffled into 141 shards (~1 GB
  each, small row groups + page index for fast random access).

## Evaluation (held-out)

Zero-shot benchmarks, kept **out of training** and shipped as separate configs. Evaluate with
per-image affine (least-squares scale + shift) alignment of predicted disparity to ground truth,
then report **AbsRel** and **&delta;<sub>1</sub>**:

| config | benchmark | notes |
|---|---|---|
| `nyuv2` | NYUv2 | indoor RGB-D; Eigen 654 test split; metric depth, capped at 10 m |
| `kitti` | KITTI | driving; Eigen test split; sparse LiDAR depth, capped at 80 m |

```python
ev = load_dataset("blanchon/dinodepth-dataset", name="nyuv2", split="test")
```

## License

Released for **non-commercial research** under **CC BY-NC-SA 4.0**. Each source retains its original
license (Virtual KITTI 2: CC BY-NC-SA 3.0; Hypersim: CC BY-SA 3.0; IRS / BlendedMVS / TartanAir per
their upstream terms) &mdash; respect the original terms of each source.

---

<sub>Data composition and recipe follow <b>AnyDepth</b> (arXiv:2601.02760).</sub>