ConAmodal

Anonymous release for double-blind review. This repository is the review copy of a dataset submitted with a paper under review. Author and affiliation information is deliberately absent. On publication the dataset moves to a permanent, DOI-bearing archive and this record is updated to point at it.

A benchmark for amodal instance segmentation on construction sites — recovering the full extent of structural elements (beams and columns, rebar, piping) that are partly hidden behind other objects.

Existing amodal benchmarks (COCOA, KINS, D2SA) cover everyday scenes: people, vehicles, household objects. Construction elements behave differently — they are long, straight, repetitive, and follow strong shape priors — and no amodal benchmark covered them. ConAmodal fills that gap.

What is in here

ConAmodal is the first two rows — 1,643 instances in 4,925 samples. The real-occlusion directory ships with them but is not part of the benchmark; see SCHEMA.md §1 for why.

Directory	Role	Occlusion	Imagery
conamodal_generation2/	train + model selection	synthetic, rate-controlled	AI-generated
conamodal_field/	test — the main benchmark	synthetic, rate-controlled	real, 5 construction sites
conamodal_field_realocc/	in-the-wild check — not part of ConAmodal	real	real, same 5 sites
splits/generation2/	frozen train/dev partition of conamodal_generation2

Every sample provides an occluded image, the ground-truth amodal mask (full shape), the ground-truth modal mask (visible part), and — where it exists — the pre-occlusion image, which serves as the RGB target for generative methods.

Three categories: structural_member, piping, rebar, with 12 subcategories. Note that id / ann_id keep the older beam_ and mep_ prefixes — read the category field, never the identifier (SCHEMA.md §3).

Design, in one paragraph

Amodal ground truth is usually guessed: an annotator looks at a partly hidden object and draws what they think is behind the occluder. ConAmodal avoids that for both of its subsets. Annotators only accepted objects that were already almost fully visible (natural occlusion ≤ 5%), so the amodal mask is close to an observed fact rather than an estimate. Occlusion is then added synthetically under a controlled rate, with occluders scaled by category-specific natural-size priors and placed depth-aware, which is what makes occlusion rate, occluder count, and fragment count usable as difficulty axes. conamodal_field_realocc is the counterweight: real occlusion, annotated by the photographer who visited the site, used to check that model rankings on the synthetic benchmark hold up.

This is a deliberate trade: ConAmodal is smaller than general-purpose amodal datasets and buys ground-truth integrity with that size.

Quick start

⚠️ Download on Linux, macOS, or WSL — not on a native Windows filesystem. Files in the synthesized subsets are named after their sample id, which joins the instance key to the occlusion band with :: (beam_IMG_8255_obj0::light.jpg). Windows does not allow : in a filename, so a checkout there fails partway. See SCHEMA.md §2.

import json
from pathlib import Path

root = Path("conamodal_field")
meta = json.loads((root / "exports" / "metadata.json").read_text())

print(meta["total_samples"], meta["categories"])
s = meta["samples"][0]
image  = root / "composites" / s["image_path"]
amodal = root / "composites" / s["amodal_mask_path"]

Before you split the data, read SCHEMA.md §4. Each instance in the synthesized sets appears three times (light / medium / heavy occlusion) sharing one ann_id. Splitting on sample id leaks the same physical object into both sides. Use the provided splits/generation2/split_manifest.json.

conamodal_field and conamodal_field_realocc are held-out evaluation sets — use them whole and do not tune on them.

Documentation

File	Contents
SCHEMA.md	layout, every metadata field, splitting rules
DATASHEET.md	motivation, collection, labeling, uses, limitations (Gebru et al. format)
LICENSE	component-by-component licensing
LICENSES-THIRD-PARTY.md	third-party material and required attribution
croissant.json	machine-readable metadata (ML Commons Croissant 1.0)
SHA256SUMS	integrity checksums

Known limitations

Please read these before drawing conclusions from the numbers.

• Scale. Per-subcategory results have low statistical power. Treat subcategory breakdowns as exploratory.
• Synthetic occlusion. The two main sets use pasted occluders. Real occlusion differs; conamodal_field_realocc exists to bound that gap, and its occlusion rates skew lighter than the controlled 1/3-1/3-1/3 split of the main benchmark.
• Real-occlusion ground truth is reconstructed, not observed, and that set is small — use it for rank-stability checks, not for reporting model scores.
• Site and schedule bias. Each site was photographed on a single day, so category coverage follows whatever construction stage was underway. One site contributes no piping imagery at all.
• Generated training imagery carries a domain gap relative to real photographs.

License

Component-specific; the combined dataset is non-commercial. See LICENSE. Third-party attribution is a license condition — see LICENSES-THIRD-PARTY.md.

Citation

[citation to be added on publication]