rayruiyang/CVBench3DTexutal

CV-Bench-3D Textual

Textual bounding-box evaluation for depth and 3D-distance reasoning.

This repository provides the GeoUG textual adaptation of the 3D subset of CV-Bench, introduced by Cambrian-1. Object references are expressed as bounding-box coordinates in the question. The repository and filename retain CVBench3DTexutal for compatibility with the GeoUG evaluation adapter.

GeoUG project · Download TSV

Contents

Property	Value
File	CVBench3DTexutal.tsv
Questions	1,200
Categories	600 Depth, 600 Distance
Answers	603 A, 597 B
Images	Encoded image bytes stored as base64 in the image column
File size	158,661,662 bytes (151.3 MiB)

Sources and adaptation

CV-Bench is the upstream benchmark. Its 3D questions are based on Omni3D; this TSV preserves the upstream provenance columns source, source_dataset, and source_filename. All 1,200 rows have source="Omni3D".

Preserved source_dataset value	Questions
Omni3D_Hypersim	400
Omni3D_SUNRGBD	400
Omni3D_nuScenes	400

The GeoUG adaptation supplies textual object references for evaluation and packages the data in a VLMEvalKit-compatible TSV. It does not claim authorship of CV-Bench, Omni3D, or the source images. Please credit the upstream CV-Bench / Cambrian-1 work and the source datasets alongside GeoUG when using this release. See the upstream data card for its description, citations, and original license information.

Download and read

import base64
import io
import pandas as pd
from PIL import Image
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="rayruiyang/CVBench3DTexutal",
    repo_type="dataset",
    filename="CVBench3DTexutal.tsv",
)
data = pd.read_csv(path, sep="\t")
row = data.iloc[0]
image = Image.open(io.BytesIO(base64.b64decode(row["image"]))).convert("RGB")
print(row["prompt"])

Use the embedded image content; no original local image directory is needed. The image_path and source_filename fields retain provenance information.

Schema and evaluation

Fields	Meaning
index	Sample identifier
image	Base64-encoded image
question, prompt	Textual question; prompt also includes answer options
A, B, answer	Option descriptions and ground-truth answer
category, split	Depth or Distance; all rows have split 3D
bbox	Reference bounding boxes; coordinates are in image pixels
source, source_dataset, source_filename, image_path	Provenance and image identifiers
target_class, target_size	Retained upstream metadata; may be empty

Only the image and question/options should be passed to the model. The answer column is for scoring. Model-specific coordinate conversion is performed by the GeoUG evaluation adapter.

From the GeoUG code checkout:

CVBENCH_3D_TEXTUAL_PATH=/absolute/path/to/CVBench3DTexutal.tsv \
MODEL=Qwen2.5-VL-7B-Instruct \
DATASETS=CVBench3DTexutal_circular_qwen25 \
WORK_DIR=outputs/vlmevalkit/cvbench_textual \
bash evaluation/vlmevalkit/run.sh --judge exact_matching

Use the coordinate-style suffix appropriate for the model. Circular evaluation runs both answer-option orders and credits a question only if both are correct. The TSV contains the 1,200 base questions; the adapter generates the rotations. Report circular accuracy separately from ordinary per-row accuracy.

File integrity

SHA256: 2dc30fe9ddf496e6e54aff14be94fa7aa7260a6dcdaf7644ae6a91ff5e707a3f

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.