3DCueBench

A balanced benchmark for depth ordering and object-to-object 3D distance reasoning.

Project page · GeoUG paper · Download parquet

3DCueBench is the spatial reasoning benchmark introduced with GeoUG: Learning and Reasoning with Geometry — How Generation Helps Spatial Understanding? Each question pairs an image with two answer options and textual point coordinates identifying the objects. A model must select A or B using their relationship in the 3D world.

At a glance

Property	Release v1.0
Questions / unique images	800 / 757
Tasks	400 depth-ordering + 400 3D-distance-ordering questions
Domains	400 indoor + 400 outdoor questions
Answer balance	400 A + 400 B
Difficulty	Four relative-gap bins; 100 questions per task per bin
Packaging	One self-contained parquet with embedded JPEG images
Hugging Face split	test
File size	115,541,247 bytes (110.2 MiB)

Depth ordering: Which of two objects is closer to the camera?

3D-distance ordering: Which of two objects is closer to a specified anchor object in 3D space?

The row-level split column identifies the task (depth or distance); it does not indicate a train/test partition. All 800 questions belong to the released test set.

Quick start

from datasets import Image, load_dataset

dataset = load_dataset("rayruiyang/3DCueBench", split="test")
dataset = dataset.cast_column("image", Image())
sample = dataset[0]

image = sample["image"]  # PIL image
print(sample["question"])  # Includes the A/B options and point coordinates

You can also read the parquet directly, without a separate image download:

import io
import pyarrow.parquet as pq
from huggingface_hub import hf_hub_download
from PIL import Image

path = hf_hub_download(
    repo_id="rayruiyang/3DCueBench",
    repo_type="dataset",
    filename="data/test-00000-of-00001.parquet",
)
row = next(pq.ParquetFile(path).iter_batches(batch_size=1)).to_pylist()[0]
image = Image.open(io.BytesIO(row["image"]["bytes"])).convert("RGB")
print(row["question"])

The image struct contains bytes and path. Use the embedded bytes; path is an image filename, not a required local file. Images are encoded as JPEG with a maximum side of 1600 pixels.

Fields

Fields	Meaning
question_id, image_id	Question identifier and image identifier
image, image_sha256	Embedded image and SHA-256 of its encoded bytes
question	Complete question, including options and point references
option_a, option_b	Object descriptions for the two options
answer	Ground-truth option, A or B
point_a, point_b	Object reference points as [x, y]
anchor, point_anchor	Anchor description and point for distance questions; null for depth questions
grid_width, grid_height	Coordinate-grid dimensions
image_width, image_height	Source-image dimensions recorded during benchmark construction; the embedded image may be resized
split	Task: depth or distance
difficulty_bin, relative_gap	Difficulty group and relative separation of the two geometric values
distance_a, distance_b	Reference geometric values used to compare the options
source, domain	Source identifier and indoor/outdoor domain
quality_severity	Construction-time quality diagnostic
shortcut_point_correct, shortcut_second_correct, shortcut_third_correct, shortcut_class_prior_correct	Construction-time shortcut-rule diagnostics; may be null

Only provide the image and question to the evaluated model. Ground-truth answers, reference distances, relative gaps, and diagnostic columns are for scoring and analysis, not model inputs.

Coordinates and difficulty

Points use an aspect-preserving grid with a long side of 1000. The origin is the top-left; x increases rightward and y increases downward. The shorter axis follows the image aspect ratio, so both axes are not necessarily 0–1000. Use the stored grid_width and grid_height when converting coordinates.

To locate a point on the decoded image, scale x by image.width / grid_width and y by image.height / grid_height.

Relative gap is abs(distance_a - distance_b) / max(distance_a, distance_b). Smaller gaps correspond to more similar geometric values.

Relative-gap bin	Depth	Distance	Total
0.10-0.16	100	100	200
0.16-0.25	100	100	200
0.25-0.40	100	100	200
0.40+	100	100	200

Evaluation protocol

Report overall A/B accuracy and breakdowns by task, difficulty, domain, and source. Missing or unparseable answers count as incorrect. Preserve the released questions, option order, and coordinate convention, or explicitly document any model-specific coordinate conversion.

In the GeoUG code checkout, save predictions as parquet, JSONL, TSV, or CSV with question_id and prediction columns, then run:

python benchmarks/3dcuebench/evaluate.py \
  --dataset /path/to/test-00000-of-00001.parquet \
  --predictions predictions.jsonl \
  --output outputs/3dcuebench/metrics.json

Each question uses one closer formulation. This release does not contain paired closer/farther questions, so closer/farther consistency is not an official v1.0 metric. Report option-order circular evaluation separately from the fixed-option-order protocol.

Sources and scope

This GeoUG benchmark combines spatial questions with images and geometry derived from the source datasets below. The source column retains the provenance identifier for each question; counts refer to this 800-question release. The upstream datasets were created by their respective authors, not by this benchmark release.

The source column contains these identifiers and question counts:

Source identifier	Original source / preparation pool	Questions
wd3d	WildDet3D	184
nuscenes	nuScenes	176
ca1m	CA-1M	136
scannetpp	ScanNet++	130
hypersim	Hypersim	128
dense	Dense-depth source pool	28
waymo	Waymo Open Dataset	16
hammer	HAMMER	2

This is a controlled evaluation set, not a training corpus or an exhaustive measure of 3D understanding. Small source-specific groups should be interpreted with their sample counts. The source images remain subject to their respective upstream dataset terms.

File integrity

The release contains one data file; no external image directory, camera table, or depth cache is needed.

data/test-00000-of-00001.parquet
SHA256: 5299fbcb85a9c0a2adf661bd89f7828b7fb332ac1c3951ba691379c78430e4b4

Publisher, permitted use, and contact

The following notice applies to this GeoUG data release:

It is published by The University of Hong Kong (HKU) and is licensed for non-commercial use only. Its use is strictly limited to academic research and educational purposes. Please contact rayyang0116@connect.hku.hk if you have any concerns regarding the data.

The original datasets are credited in the sources section. This release does not transfer ownership of upstream images or annotations; their original licenses, attribution requirements, and terms of use remain applicable.