blanchon/dinoflow-dataset

DinoFlow

A harmonized, pre-shuffled corpus of 61,514 (frame1, frame2, flow) pairs for training two-frame optical-flow models on a frozen DINOv3 backbone. The synthetic training sources are decoded to a single dense-flow 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
image1	Image	frame t (PNG, lossless RGB)
image2	Image	frame t+1 (PNG, lossless RGB)
flow	binary	float16 LE, (H, W, 2) row-major, (u, v) pixels (frame1 → frame2)
valid	Image	uint8 1-channel PNG, 255 = valid, 0 = invalid/occluded

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

ds = load_dataset("blanchon/dinoflow-dataset", split="train", streaming=True)
row = next(iter(ds))
frame1 = row["image1"]                                              # PIL RGB
frame2 = row["image2"]                                              # PIL RGB
H, W = row["height"], row["width"]
flow = np.frombuffer(row["flow"], "<f2").reshape(H, W, 2)           # (u, v) pixels
valid = np.asarray(row["valid"]) > 0                               # bool (H, W)

Composition

source	pairs	domain	flow
FlyingThings3D_subset	39,282	objects in 3D scenes (synthetic)	dense
FlyingChairs	22,232	chairs over backgrounds (synthetic)	dense
total	61,514

Harmonization

Every source is decoded to dense flow (u, v) in pixels (frame1 → frame2), stored as compact little-endian float16 (near-lossless for magnitudes < ~1000 px). The synthetic flow is dense; pixels with |u| or |v| ≥ 1000 px are marked invalid in valid (the FlowNet/RAFT convention). Frames are stored losslessly as PNG at native resolution; resize/crop happens in the training loader.

Splits

• train — the full 61,514-pair corpus, globally shuffled into 121 shards (~1 GB each, small row groups + page index for fast random access).

The train split is the full FlowNet/RAFT C→T training corpus: FlyingChairs (the C stage) and FlyingThings3D_subset (the T stage), decoded to one flow convention and globally shuffled together, so any shard mixes both sources.

Evaluation (held-out)

Standard optical-flow benchmarks, kept out of training and shipped as separate configs. Report EPE (mean end-point error over valid pixels) and, on KITTI, Fl-all (outlier rate):

config	benchmark	notes
sintel-clean	MPI-Sintel (clean)	movie-grade synthetic; dense flow; clean render pass
sintel-final	MPI-Sintel (final)	as clean, with motion blur + atmospherics
kitti2015	KITTI-2015 flow	driving; 200 pairs; sparse LiDAR-derived flow + valid mask

ev = load_dataset("blanchon/dinoflow-dataset", name="sintel-clean", split="test")

License

Released for non-commercial research under CC BY-NC-SA 4.0. Each source retains its original license — respect the original terms of each source (FlyingChairs / FlyingThings3D per their upstream LMB terms; MPI-Sintel and KITTI-2015 per their respective terms).

---

_{Recipe mirrors DinoDepth / AnyDepth (arXiv:2601.02760), for optical flow.}