ARROW

Agentic Reconstruction from Real-world Observations in the Wild is a benchmarkable dataset of particular physical objects: photographs taken in ordinary surroundings, descriptions at three levels of detail, reference scans and measured mass. Code, validator and data dictionary.

An asset can be visually convincing while being wrong about the object it represents. A handle may be closed, a cavity filled, a narrow contact surface displaced or the entire object reconstructed at the wrong scale. For instance reconstruction, plausibility is not sufficient: the target is the particular object in the photographs, not an attractive substitute.

ARROW makes the reconstruction recipe the experimental unit. Compare complete pipelines on the same observations, or hold a pipeline fixed and replace one component. The reference geometry and mass provide an independent target for measuring the resulting object. Record changes to prompts, adapters and preprocessing alongside component substitutions so that the comparison has a precise meaning.

Just some examples from the dataset

The overview table contains every case. These examples pair an in-the-wild photograph with six orthographic views of its reference mesh; front/back/left/right/top/bottom refer to the scanner axes.

Ice-cream bowl

Field	Value
Long description	This is one open ice-cream bowl made from vacuum double-walled 316 stainless steel. It has a wide circular rim and a deep, continuously curved interior suitable for an individual serving. The exterior is smoothly rounded and narrows towards a small flat base. No handles, lid, decoration or printed markings are visible. The brushed silver surface retains fine scratches and reflections from the surrounding cupboard.
Medium description	A round stainless-steel bowl has a wide circular rim, deep curved interior and smoothly rounded exterior. Its wall narrows towards a small flat base, with no handles, lid or decoration. The brushed reflective finish shows fine scratches and warm reflections from the surroundings.
Short description	A double-walled stainless-steel ice-cream bowl.
GLB	View GLB
Full metric mesh	Download mesh
Subset	ARROW / household_items
Date acquired	2026-09-06
Acquired by	@chrisvoncsefalvay

Coke can

Field	Value
Long description	This is one tall, slim Diet Coke drinks can made from aluminium. Its body is cylindrical, with shallow curved shoulders, rolled top and bottom rims, and an open stay-tab lid. The silver surface carries large red vertical lettering and smaller red text near the base. Window light creates strong reflections across the metal. Minor dents and handling marks remain visible.
Medium description	A tall, slim aluminium drinks can has straight cylindrical sides, shallow shoulders and rolled rims. Its silver printed body carries large red Diet Coke lettering and small red text near the base. The stay-tab lid is open, and the reflective surface shows light dents and handling marks.
Short description	A silver Diet Coke aluminium drinks can.
GLB	View GLB
Full metric mesh	Download mesh
Subset	ARROW / containers_packaging
Date acquired	2026-09-07
Acquired by	@chrisvoncsefalvay

Mug

Field	Value
Long description	This is one ceramic souvenir mug from Sanderling Resort in Duck, North Carolina. Its body is rounded and slightly waisted, widening towards a flared rim and narrowing to a small footed base. A broad loop handle joins the upper and lower body. The exterior has a mottled blue-green glaze with warm brown edging. A raised cream plaque bears a bird emblem, the resort name and its location, and the mug shows the irregular finish typical of decorative glazed pottery.
Medium description	A ceramic souvenir mug has a rounded, slightly waisted body, flared rim and small footed base. Its broad loop handle curves from the upper wall to the lower body. The mottled blue-green glaze has brown edging and a raised cream Sanderling Resort plaque.
Short description	A glazed Sanderling Resort ceramic souvenir mug.
GLB	View GLB
Full metric mesh	Download mesh
Subset	ARROW / household_items
Date acquired	2026-09-06
Acquired by	@chrisvoncsefalvay

What loading returns

overview is the compact browsing table. cases and case_inputs specify the evidence for each case; observations and assets resolve that evidence to files and checksums. descriptions retains the collector's wording and the three reviewed observer descriptions, including attribution. ground_truth_records links the reference geometry and physical_properties stores measured values with their units. The data dictionary describes every field.

development contains cases with verified geometry. candidate contains additional photographed and weighed objects whose surface issues are recorded by geometry_incomplete. Collection state, reference maturity, reference access and benchmark partition are separate columns. A null physical value means that the corresponding measurement is unavailable.

Load a pinned release

from datasets import load_dataset

repo_id = "chrisvoncsefalvay/ARROW"
revision = "rel_01M23Z8GJMWPSZMBKKWGZ8H8MP"
objects = load_dataset(repo_id, "overview", split="development", revision=revision)
object_row = objects[0]
object_row["six_view"].show()
print(object_row["description_medium"], object_row["mass_g"])

revision names the immutable release tag. The Parquet tables load with standard Hugging Face tooling; images, GLBs and scans are separate checksum-addressed files.

Feed observations to a pipeline

Install the ARROW tools, then select the three-sentence observer description and the case's in-the-wild photograph:

from arrow_bench.client import load_case, write_submission

evidence = load_case(
    object_row["case_id"], revision=revision,
    description_tier="structured_observer",
)
# Your pipeline consumes evidence["description"] and evidence["photo_paths"].
# Save its generated mesh as generated.glb, then register the actual output:
write_submission(
    "generated.glb", case_id=evidence["case_id"],
    output="submission", declared_length_unit="m",
)

The helper returns the selected description ID and photograph hashes. Retain this metadata with the pipeline's run record. Using the same selected description across pipelines fixes their text input. The original collector wording remains available through case_inputs.

Compare the output with the reference

from arrow_bench.client import download_release

reference = download_release(revision=revision, local_dir="reference")

arrow-bench evaluate ./submission --dataset-release ./reference \
  --suite geometry-v0 --output ./evaluation

The validator resolves the matching reference mesh by case ID and reports surface accuracy, completeness, normal agreement, topology and metric size. Mass and density errors are calculated when the submission provides compatible predictions. Rigid and similarity registration have separate results and saved transforms. CUDA runs the surface-distance queries; W&B records the evaluation offline.

The identity recipe copies the reference into a submission and checks the evaluator's exact-match behaviour. Its recipe ID is arrow.identity-oracle.

Release inventory

Release rel_01M23Z8GJMWPSZMBKKWGZ8H8MP · schema arrow.dataset/1.0.

Table	Rows
assets	50
capture_sessions	5
case_inputs	17
cases	5
descriptions	20
ground_truth_records	10
identity_claims	7
object_states	5
observations	5
overview	5
physical_properties	5
release_manifest	5
specimens	5

Citation

@misc{arrow2026dataset,
  author = {von Csefalvay, Chris},
  title = {ARROW: Agentic Reconstruction from Real-world Observations in the Wild},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {Dataset},
  url = {https://huggingface.co/datasets/chrisvoncsefalvay/ARROW},
  version = {rel_01M23Z8GJMWPSZMBKKWGZ8H8MP}
}