DannHiroaki/VOSMA-Dataset

VOSMA Real Rectangle Datasets

Construction pipeline for three real-data rectangle relations for area-weighted and IoU-weighted spatial join sampling. Release v0.2.0 contains completed DocLayNet, MOT20, and COCO/Sama-COCO datasets. Each retained source annotation or exported prediction becomes one rectangle with unit record weight. This release contains integer arrays and provenance metadata; images, segmentation masks, algorithm indexes, and materialized join pairs are not included.

Dataset directory	R1	R2	Retained R1 records	Retained R2 records	Groups
doclaynet	Official test-page ground truth	Fixed Aryn model predictions	66,531	50,084	4,999
mot20	Valid pedestrian ground truth	Official detections	1,134,614	661,143	8,931
coco_sama	COCO 2017 non-crowd instances	Sama-COCO non-crowd instances	886,282	1,068,028	386,072

Counts refer to this processed release. Empty sides and groups remain represented according to the group universes below. The per-dataset dataset.json and validation.json provide filtering counts and validation results.

Release status

All 4,999 selected official DocLayNet test pages completed fixed-model inference and the independent image/provenance audit. The three datasets passed record-level geometry and group-isolation validation.

Completed dataset	Positive-area record pairs	Groups with positive mass	Groups with zero mass
DocLayNet	52,670	4,876	123
MOT20	3,176,148	8,931	0
COCO/Sama-COCO	1,508,091	322,274	63,798

The pair counts are validation statistics. No pair list is distributed. COCO has two zero-height non-crowd boxes; Sama has eight zero-width/height non-crowd boxes. These ten source records are itemized in coco_sama/independent-audit.json. All 119 train and 5 validation Sama shards passed an independent manifest, identity, grouping, and row-count audit. Out-of-bounds diagnostics use quantized local coordinates and the source's pixel origin.

Builder: DANNHIROAKI/VOSMA-Dataset-Build. Built arrays: DannHiroaki/VOSMA-Dataset. See BUILD_INFO.json for the construction commit and environment.

Source selection and grouping

DocLayNet. R1 contains all original precedence-0 ground-truth boxes on all 4,999 official test pages from the DocLayNet 1.0.0 Core archive, using COCO/test.json. R2 contains the exported predictions of the fixed Aryn/deformable-detr-DocLayNet model. A group is (split, source_image_id). The physical key (doc_category, collection, doc_name, page_no) is unique across selected pages; official PNG filename, dimensions, source identity, and per-image bytes are checked. All selected pages remain groups, including successful pages with zero predictions. Missing, failed, or duplicate inference pages stop the build. All classes remain in their page groups, so cross-class intersections are included. Predictions are not matched to GT or selected using GT.

The released model run uses CPU float32, batch size 1, and four PyTorch threads per worker. It began with eight workers and resumed with sixteen workers under the same frozen numerical protocol; completed page results were reused only after protocol and image-hash checks. The reproduction command uses sixteen workers. The fixed processor produces [1,3,800,800] model inputs and restores predicted xyxy boxes to each original PNG's dimensions. Native postprocessing takes the top 100 scores over the flattened 200-query × 12-class sigmoid output. The mutually exclusive exclusions are nonfinite scores, finite scores at most 0.7, and remaining class-0 (N/A) candidates, in that order. All remaining records are exported. There is no NMS, clipping, deduplication, GT-based threshold tuning, or class-pair filtering. Original top-100 rank, query index, label, score, and PNG-coordinate endpoints are retained in metadata. Geometry exclusions are applied subsequently by the common dataset conversion.

The checkpoint revision is d5503a90ae08dd43565de6984a5dd7924cad2400 and its model-card license is Apache-2.0. The model's training overlap with these DocLayNet test pages is unknown. These relations measure spatial-join sampling on GT and frozen predictions; they do not establish an unseen-test-set detection result. Model files, page selection, protocol, predictions, and image hashes are recorded in provenance; doclaynet/inference-audit.json reports the independent all-page PNG, provenance, and prediction-accounting audit. See doclaynet/README.md for the completed counts and hash links.

MOT20. The source is MOT20Labels.zip. Only training sequences MOT20-01, MOT20-02, MOT20-03, and MOT20-05 are used. A group is (sequence, frame); all 8,931 frames are retained. R1 keeps GT rows with valid=1 and class=1, without a visibility threshold. R2 keeps official detections with a finite score and valid geometry, without a score threshold. Detection scores, track IDs, and visibility are metadata; they never become sampling weights. The original MOT coordinate convention is preserved before quantization, with no one-pixel origin adjustment. Official benchmark and downloads.

COCO/Sama-COCO. R1 uses the COCO 2017 train/validation instance annotations; R2 uses the official Sama train and Sama validation annotations. Images are paired by source file identity within the original split, with dimension checks. Category names establish the mapping between the 80 classes. A group is (split, image, category) from the union of category mentions on either side before filtering. Both sides require iscrowd=0. Annotation-free images remain in images.parquet; unmentioned image-category combinations are not synthesized as groups. Sama reannotation began with COCO annotations, so this is a comparison of annotation versions, not a claim of independent blind annotation. Sama's description of its labeling process.

Geometry contract

For each source bbox=[x,y,w,h], construct the endpoints exactly from the original decimal values, then compute:

Q(z) = nearest integer to 1000*z, with exact ties rounded to even
(x0,y0,x1,y1) = (Q(x), Q(y), Q(x+w), Q(y+h))

Raw annotation decimals are parsed as Decimal, and source endpoint arithmetic is exact. The DocLayNet model itself intentionally runs in float32. Its restored PNG-coordinate xyxy outputs are serialized with Python float round-trip precision, read back as exact decimals, and converted to xywh with exact endpoint differences before Q1000 quantization. Dataset conversion performs no additional float32 coercion after the model output. Rectangles are half-open: [x0,x1) × [y0,y1). A join pair requires strictly positive intersection width and height; boundary contact alone is not an intersection.

Malformed or non-finite boxes, nonpositive source widths/heights, and rectangles that become degenerate after quantization are excluded and counted. Finite negative coordinates and boxes outside the declared image bounds are retained. Coordinates are not clipped or resized during dataset conversion. Representability failures stop the build rather than silently wrapping integers.

After quantization, each group receives one shared integer translation for both R1 and R2. Its minimum retained y-coordinate becomes 1. Groups occupy disjoint x intervals with a one-unit gap, using the union of retained rectangle extents on both sides. The translation and local bounds are stored in groups.parquet. Translation preserves the quantized within-group geometry, areas, and IoUs while preventing intersections between different groups. Quantization itself may change the original continuous geometry slightly.

Every output weight is 1. Area sampling uses the quantized intersection area; IoU sampling uses intersection area divided by union area. No object matching, IoU threshold, NMS, deduplication, bbox merging, or annotation-count balancing is performed. Equal numbers of samples per group do not implement global mass-weighted sampling. Algorithm preprocessing remains part of measured algorithm cost. Int64 coordinate storage does not guarantee that area or cumulative-mass intermediates fit int64. Equal rectangles remain distinct records when they originate from distinct source annotation or prediction records.

Files and schema

Each of the three completed dataset directories contains the following files.

File	Contents
R1.npy, R2.npy	NumPy arrays with shape (N,7), little-endian signed int64, no pickle
R1_metadata.parquet, R2_metadata.parquet	One row per retained rectangle, aligned with array row order
groups.parquet	Group identity, image dimensions, common translation, bounds, contiguous array slices, empty-side state, and validated overlap counts
dataset.json	Schema version, relation definitions, counts, filtering statistics, and numerical contract
validation.json	Validation results and clearly scoped geometric spot checks
exclusions.jsonl.gz	Rejected source records and exclusion reasons
SHA256SUMS	Checksums for the processed files

Dataset-specific pairing reports, category mappings, and images.parquet are included where applicable. Each dataset's sources/*.json records source URLs, versions, retrieval metadata, archive members, and SHA-256 checksums of the extracted original annotation files. Source fingerprints identify the exact input bytes used for the release.

The seven array columns are:

record_id, group_id, x0, y0, x1, y1, weight

record_id is a zero-based row number unique within its relation; (relation, record_id) is the full generated record identity. It is not the source annotation ID. group_id is shared across the two relations. Coordinates are quantized and translated; weight is always 1.

The source record position is a 1-based CSV line number for MOT20 and a 0-based annotations-array index for JSON sources. Metadata includes source file, record position, annotation/image/category IDs, the original bbox decimal values, quantized local endpoints, and source-specific attributes. In groups.parquet, r1_start/r1_count and r2_start/r2_count identify contiguous row slices. state is one of both_nonempty, r1_only, r2_only, or both_empty. A group with two nonempty sides can still have zero positive-area join pairs.

Load a frozen release

Use an immutable Hugging Face commit SHA rather than a moving branch. Install numpy, pyarrow, and huggingface_hub, then:

from pathlib import Path
import json
import numpy as np
import pyarrow.parquet as pq
from huggingface_hub import snapshot_download, HfApi

revision = HfApi().dataset_info("DannHiroaki/VOSMA-Dataset", revision="v0.2.0").sha
print("Record this immutable revision with your experiment:", revision)
dataset = "doclaynet"  # also available: mot20, coco_sama
root = Path(snapshot_download(
    repo_id="DannHiroaki/VOSMA-Dataset",
    repo_type="dataset",
    revision=revision,
    allow_patterns=[f"{dataset}/*"],
)) / dataset

r1 = np.load(root / "R1.npy", mmap_mode="r", allow_pickle=False)
r2 = np.load(root / "R2.npy", mmap_mode="r", allow_pickle=False)
groups = pq.read_table(root / "groups.parquet")
info = json.loads((root / "dataset.json").read_text())
assert info["schema_version"] == "vosma-real-rectangles-v1"
assert r1.shape[1] == r2.shape[1] == 7

# Read one group's original row slices, including an empty side if present.
g = groups.slice(0, 1).to_pylist()[0]
a = r1[g["r1_start"]:g["r1_start"] + g["r1_count"]]
b = r2[g["r2_start"]:g["r2_start"] + g["r2_count"]]
print(r1.shape, r2.shape, groups.num_rows, a.shape, b.shape)

The arrays can be used directly as benchmark inputs. Keep download, model inference, source conversion, and file-loading costs separate from algorithm measurements. Algorithm preprocessing remains part of measured algorithm cost. Report the release revision and requested sample count with results.

Rebuild from original annotations

From the builder repository root, with Python 3.12 and a C++17 compiler available:

python3 -m venv .venv
.venv/bin/python -m pip install -r requirements.txt -r requirements-inference.txt
.venv/bin/python fetch_sources.py --raw-root raw --sources doclaynet mot20 coco2017 sama_train sama_val
.venv/bin/python fetch_doclaynet_model.py --output-root runtime/doclaynet
.venv/bin/python fetch_doclaynet_images.py --raw-root raw --output-root runtime/doclaynet --splits test --workers 16
.venv/bin/python infer_doclaynet.py --pages runtime/doclaynet/pages.json --images runtime/doclaynet/images --model runtime/doclaynet/model --model-manifest runtime/doclaynet/model-manifest.json --output-root runtime/doclaynet/predictions --workers 16 --threads 4
.venv/bin/python build.py --raw-root raw --output-root data --datasets doclaynet mot20 coco_sama --doc-pages runtime/doclaynet/pages.json --doc-predictions runtime/doclaynet/predictions/predictions.jsonl.gz --doc-protocol runtime/doclaynet/predictions/protocol.json
.venv/bin/python -m unittest discover -s tests
.venv/bin/python validate.py --data-root data --work-root runtime/validation --datasets doclaynet mot20 coco_sama
.venv/bin/python audit_doclaynet_inference.py --pages runtime/doclaynet/pages.json --images-manifest runtime/doclaynet/images-manifest.json --predictions runtime/doclaynet/predictions/predictions.jsonl.gz --protocol runtime/doclaynet/predictions/protocol.json --images runtime/doclaynet/images --output data/doclaynet/inference-audit.json
.venv/bin/python audit_doclaynet_relations.py --raw-root raw --data-root data --pages runtime/doclaynet/pages.json --predictions runtime/doclaynet/predictions/predictions.jsonl.gz --output data/doclaynet/independent-audit.json
.venv/bin/python independent_coco_audit.py --raw-root raw --data-root data --output data/coco_sama/independent-audit.json

Retain the independent image/protocol/prediction audit as doclaynet/inference-audit.json and the full relation audit as doclaynet/independent-audit.json. Both image fetching and inference began with eight workers and resumed with sixteen; inference kept four PyTorch threads per worker. Construction concurrency is recorded separately from the frozen numerical protocol. The original construction's run-history.json records the complete run; the inference summary's elapsed time covers only its resumed phase.

The annotation fetcher uses HTTP byte ranges and ZIP/ZIP64 member extraction. It enforces sources.lock.json, checks archive ETag/size and the exact selected member set, verifies ZIP CRC32 when reusing files, and pins each extracted member's SHA-256. The DocLayNet image fetcher additionally obtains only the selected official test PNGs needed for model inference. These images are build inputs and are not redistributed in the processed dataset. Model files are pinned by revision, size, and SHA-256; inference verifies the model manifest before loading.

The adapter checks each result's exact protocol SHA-256, the protocol's page-manifest SHA-256, fixed postprocessing, page completeness, dimensions, and the original category map. The independent inference audit reads all selected PNG bytes and checks image hashes, CRCs, dimensions, prediction identity, and provenance. The separate relation audit checks full record membership against the selected original GT and frozen predictions, including geometry filtering, metadata, and group assignment. The validator checks output records, IDs, group slices, unit weights, common translations, actual slab containment, and group separation, and counts positive-area overlaps without saving join pairs. Geometric spot checks are diagnostics, not full-pair population estimates. The independent inference audit checks the protocol's recorded model manifest; it does not itself reread checkpoint weights.

Licenses

Licenses apply to each dataset portion separately; this collection does not replace them with one blanket license.

Portion	Source license	Conditions carried with this release
DocLayNet-derived data	CDLA-Permissive-1.0	Preserve source attribution and license access; identify the modifications described above.
Aryn source model	Apache-2.0 model card	Model weights are fetched as build inputs; their pinned revision and license remain recorded in provenance.
MOT20-derived data	CC BY-NC-SA 3.0	Attribution, noncommercial use, and ShareAlike conditions apply.
COCO/Sama-COCO-derived annotation data	CC BY 4.0	Preserve attribution and license access, and indicate modifications.

Credit the original DocLayNet, MOT20, COCO, and Sama-COCO creators when using their portions. This derivative changes annotation selection, numerical representation, and group placement as documented above. Source image licenses are separate; source images are not redistributed here. See the COCO terms and Sama dataset information for the original resources.

Cite the sources

• Pfitzmann et al. (2022), DocLayNet: A Large Human-Annotated Dataset for Document-Layout Analysis.
• Dendorfer et al. (2020), MOT20: A benchmark for multi object tracking in crowded scenes.
• Lin et al. (2014), Microsoft COCO: Common Objects in Context.
• Zimmermann et al. (2023), Benchmarking a Benchmark: How Reliable is MS-COCO?.

Also identify this dataset repository and the exact release commit SHA in experiment reports. Model predictions versus GT, detection versus GT, and annotation-version comparisons have different provenance and should be reported separately.