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

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

Harmonization

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

• Hypersim — ray distance → planar Z (per-pixel, focal 886.81 px).
• Virtual KITTI 2 — native depth / 100 (cm → m).
• IRS — disparity → depth (48 / disparity).
• TartanAir — native metres; sky masked.
• BlendedMVS — 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 — a scale-and-shift-invariant objective absorbs it, so the sources mix directly.

Splits

• train — 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 δ₁:

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) — respect the original terms of each source.

_{Data composition and recipe follow AnyDepth (arXiv:2601.02760).}