RA3D collision and measurement research data

Data and intermediate results for Holographic Droplet Collision Measurement, by Dai Nakai. The RA3D Python repository provides the algorithms, selective downloader, training, inference, and offline visual review.

This release contains six synthetic scenes, each with 6,000 synchronized frame pairs and 2,000 collision events. Seeds 260740–260744 form the development/training split. Seed 260745 is the terminal validation scene and is excluded from every fitted model. All collision-model fits use synthetic data only.

Download only what you need

Use ra3d list-data to inspect the exact sizes. Select tiers and scenes explicitly:

ra3d download --tier features --tier labels --split train --output data
ra3d download --tier tracking --scene 260745 --output data
ra3d download --tier holograms --scene 260745 --extract --output data

Images are never added to a feature or tracking download. Holograms are split into 500-frame-pair tar shards, with exact frame bounds in manifest.json. Each artifact has a byte count and SHA-256; the repository pins an immutable dataset revision.

Tier	Content
labels	Complete simulated events, particle states, initial particles and trajectories
detections	Raw detections and reconstructed 3D particles
tracking	Raw and smoothed trajectories; track-to-simulator identity mappings
candidates	Permissive collision candidate pools
features	Gate training tables, 77 trajectory features, static scores and paper92 scoring unions
image-complete	All 92 features for every row in the paper's final scoring union
scores	Cross-fitted and nested cross-fitted classifier scores
predictions	Frozen 3D and depth-free 2D event predictions
holograms	All paired background-normalized holograms, optical calibration and frame manifests
models	Verified collision model weights in skops/safetensors formats
baseline2d	Independent depth-free baseline inputs and intermediate scores
demo	Small self-contained event review example
holod3-models	Four upstream particle detector/depth/diameter checkpoints and notices
holod3-initializers	Two detector retraining initializers and AGPL notices
holod3-training	Five original upstream training bundles with internal validation/test splits
holod3-evaluation	Twelve paired benchmark frames, calibration and particle truth

The available files and tiers are authoritative in the manifest. Published experimental review counts do not establish exhaustive experimental event recall; this dataset's exhaustive event truth is synthetic.

Feature missingness

paper92.parquet retains the original image acquisition policy: images are evaluated for the top 25% of the static shortlist. The scoring union also contains the structural top 20 per vanishing track. Non-image rows have missing raw image measurements and zero derived image flags, exactly as in the paper.

image_complete92.parquet adds image features for every row of that final scoring union. It preserves paper_image_enriched, and adds image_features_computed. It is an additional reusable dataset; replacing paper92 by this table changes the learning problem. The wide pool and static shortlist can contain candidates outside the final scoring union; their images are not precomputed in this tier, and can be computed with ra3d features --kind image.

Coordinates and labels

Track frame 0 corresponds to hologram 000001.png. Trajectories store x/y in pixels, z in slice units, diameter in pixels, and velocities per frame. The paper uses 10 µm/pixel, 100 µm/slice and 4,000 frames/s. Collision candidate geometry and optical reconstruction have different depth origins; follow the explicit acquisition manifest and coordinate documentation.

Raw candidate identity is (scene, vanished_track_id, survivor_track_id, raw candidate frame). Coordinate correction can change the displayed frame. Never join reviews or labels using the corrected frame.

Reference evaluation

The projected-XY metric uses maximum one-to-one matching, ±5 frames and 350 µm in XY, with no Z gate. The released predictions give:

Method	Development nested OOF AP	Terminal AP
RA3D 3D	0.9110463792	0.9171254959
Depth-free 2D	0.8889090248	0.8993428951

GPU hardware and numerical libraries can affect attention retraining. The release distinguishes frozen predictions, verified deployment weights, and newly fitted results; it does not claim byte-identical neural training across all hardware.

License and attribution

The original RA3D synthetic research data in this dataset are released by Dai Nakai under CC BY 4.0. Cite the dataset revision, RA3D repository version, and associated paper. RA3D-specific collision model artifacts are covered by the MIT license supplied in their model directory. Third-party software and any separately supplied upstream model assets retain their own licenses.

Public HoloD3 measurement assets

RA3D 0.1.1 adds six data bundles and six production/initializer checkpoints, totaling 842,426,281 bytes before small notices. Public redistribution of these specific assets was explicitly authorized by Dai Nakai. They preserve the original source revisions and SHA-256 values, and are separate from the six-scene collision corpus above. Use RA3D 0.1.1 or newer for the declared extraction layout.

ra3d download --tier holod3-models --output data
ra3d download --tier holod3-training --tier holod3-evaluation --extract --output data
ra3d download --tier holod3-initializers --output data

The upstream bundles contain their original train/validation/test splits. RA3D's collision scene/split filters do not subdivide them. Each archive extracts below upstream/holod3/data/downloaded/; weights are below upstream/holod3/models/. The upstream guide shows how to clone the matching source at this destination and run the complete training ledger without authentication.

The dataset has component-specific licenses: project-owned HoloD3 data use CC BY 4.0; the three depth/diameter checkpoints use MIT; the YOLO detector and both initializers retain Ultralytics AGPL-3.0. Read LICENSES.md and the notices supplied with each tier. The dataset-wide metadata does not relicense third-party assets.