README.md

metadata

pretty_name: CDBench
license: cc-by-nc-sa-4.0
language:
  - en
task_categories:
  - visual-question-answering
  - question-answering
size_categories:
  - 10K<n<100K
tags:
  - change-detection
  - multimodal
  - vision-language
  - benchmark
  - remote-sensing
  - industrial-inspection
  - image
  - text
configs:
  - config_name: viewer
    data_files:
      - split: train
        path: questions.csv

CDBench: A Comprehensive Multimodal Dataset and Evaluation Benchmark for General Change Detection

16,086 image pairs · 73,122 multiple-choice QA items · 7 tasks · 3 domains

Project website | Dataset files | Questions and feedback

CDBench evaluates how multimodal large language models (MLLMs) understand images and reason about changes across remote sensing, industrial inspection, and commodity products. It connects image content recognition with change discrimination, regional localization, semantic classification, description, and plausible cause inference through a shared multiple-choice question (MCQ) format.

The accompanying ChangeAgent framework combines an MLLM with expert visual guidance and retrieval-augmented generation. In the reported evaluation, it achieves 78.25% average accuracy across Q1–Q7, compared with 71.19% for Gemini 1.5 Pro and 70.68% for GPT-4o.

Dataset at a glance

The counts below are computed directly from the released questions.csv. An image pair is a unique (target_image, reference_image) tuple; a QA item is one CSV row.

Domain	Source dataset	Image pairs	QA items	Task coverage
Remote sensing	CDD	2,343	9,468	Q1–Q4, Q6–Q7
Remote sensing	LEVIR-CD	3,093	12,444	Q1–Q4, Q6–Q7
Remote sensing	SYSU-CD	2,316	9,409	Q1–Q4, Q6–Q7
Industrial inspection	MVTec-AD	1,685	9,442	Q1–Q7
Industrial inspection	MVTec-LOCO	1,557	8,016	Q1–Q7
Industrial inspection	VisA	2,132	10,189	Q1–Q7
Commodity products	GoodsAD	2,960	14,154	Q1–Q7
Total	7 source datasets	16,086	73,122	7 tasks overall

Not every image pair has all seven questions. The released remote-sensing subsets contain no Q5 items, and coverage varies across tasks and samples. The totals describe the released annotations, not the sizes of the original source datasets.

Seven evaluation tasks

ID	Task	What the model selects	QA items
Q1	Image Content Classification	The primary scene type or dominant content category of one image.	16,075
Q2	Image Content Description	The description that best matches objects, attributes, composition, and spatial relations in one image.	8,995
Q3	Change Discrimination	Whether a meaningful change is present relative to the reference image.	16,038
Q4	Change Localization	A regional or semantic location where the change occurs.	8,991
Q5	Semantic Change Classification	The category of the observed change or defect.	5,034
Q6	Change Description	The description that best summarizes the observed differences.	8,995
Q7	Change Cause Inference	A plausible cause or underlying reason for the change.	8,994

Q1–Q2 assess single-image understanding; Q3–Q7 assess paired-image change analysis. All seven tasks use MCQ answer selection, including Q2 and Q6. Q4 evaluates location-related answers, not pixel-level segmentation. Q7 evaluates plausible interpretation from available evidence, not experimentally established causality.

Representative MCQs. Shaded options mark annotated answers. Panels (a)–(b) use single images; in (c)–(f), the reference/earlier image appears above the target/later image.

Dataset construction

1. Integrate visual sources. The benchmark combines seven datasets across three domains. Where a source lacks paired references, nearest-neighbor retrieval selects relevant normal examples. For remote-sensing images with limited prior information, CLIP-based scene classification supplies additional context.
2. Draft and refine questions across models. Available images, masks, labels, descriptions, and domain knowledge support question construction. GPT-4o drafts questions that Claude reviews and refines, and the models also exchange these roles.
3. Verify with two experts. Experts inspect and correct QA items for question relevance, answer validity, and ambiguity through a collaborative annotation platform.

This procedure is designed to reduce generator-specific bias and ambiguous answers. Masks and labels used to ground annotation are not automatically inputs to the evaluated MLLMs.

CDBench construction and task organization. The verification cards illustrate a correct answer and an intentionally incorrect alternative; they are examples of the checks, not historical rejection records.

Repository files and availability

The current release contains questions.csv, six image archives, and documentation images in assets/:

questions.csv
CDD.zip
LEVIR-CD.zip
SYSU-CD.zip
MVTec-AD.zip
MVTec-LOCO.zip
VisA.zip
assets/
README.md

GoodsAD availability: questions.csv includes 14,154 GoodsAD QA items covering 2,960 image pairs, but this repository does not currently contain a GoodsAD.zip archive. Downloading the six available archives therefore does not provide all images referenced by the CSV. For the matching GoodsAD files, contact the maintainers through the Community tab. If evaluating only available subsets, report that subset explicitly.

The Hugging Face viewer exposes the annotations as the viewer configuration with a split named train. This is the current file-loading configuration; it does not define an official training/evaluation partition. The reported benchmark evaluates models without CDBench-specific fine-tuning.

Annotation schema

Field	Type	Meaning
target_image	string	Relative path to the target image.
question_num	integer	Task identifier, from 1 to 7.
question	string	Question stem in English.
options	string	Lettered answer choices, usually separated by newlines.
answer	string	Correct option letter: A, B, C, or D.
reference_image	string	Relative path to the reference image.
mask	string or empty	Relative path to an available change/defect mask; some entries are empty.

Image fields contain paths, not embedded image bytes. The CSV contains quoted multiline option strings, so use a CSV parser rather than splitting the file by line. An empty mask field means no mask path is supplied for that row; it should not be treated as a negative-change label.

One released Q4 record is shown below. Newlines in options are escaped here for readability.

{
  "target_image": "CDD/B/10_154.png",
  "question_num": 4,
  "question": "Where have the most significant changes occurred between Time Points A and B?",
  "options": "A. In the agricultural regions;\nB. Near large water bodies.;\nC. In the downtown urban area;\nD. Along the roadway infrastructure",
  "answer": "D",
  "reference_image": "CDD/A/10_154.png",
  "mask": "CDD/label/10_154.png"
}

Quick start

Load the annotations

pip install datasets huggingface_hub pillow

from datasets import load_dataset

qa = load_dataset("MM-CD/CDBench", name="viewer", split="train")
print(qa.num_rows)  # 73122 in this release
print(qa[0]["question"])
print(qa[0]["options"])

# Example: select all regional-localization questions.
q4 = qa.filter(lambda row: row["question_num"] == 4)

Loading this CSV configuration retrieves the annotations; it does not download or decode the image archives.

Download image archives

Download only the source datasets you need. For example:

from huggingface_hub import hf_hub_download

archive = hf_hub_download(
    repo_id="MM-CD/CDBench",
    repo_type="dataset",
    filename="CDD.zip",
    local_dir="./CDBench",
)
print(archive)

Repeat with another filename from the repository listing as needed. Each archive is several gigabytes; the six archives total approximately 25.24 GB before extraction.

Arrange extracted files to match the CSV paths. Archive directory layouts differ: the remote-sensing archives start with an mmcd_set/ wrapper, while the industrial archives start with object-category folders. Inspect the extracted tree and place each source under its dataset-name directory, avoiding extra wrapper levels. For example, the resulting image root should contain:

images/
├── CDD/
│   ├── A/10_154.png
│   ├── B/10_154.png
│   └── label/10_154.png
├── LEVIR-CD/
├── SYSU-CD/
├── MVTec-AD/
├── MVTec-LOCO/
└── VisA/

After arranging the files, open a pair using the paths in its record:

from pathlib import Path
from PIL import Image

image_root = Path("./CDBench/images")
row = qa[0]

with Image.open(image_root / row["reference_image"]) as image:
    reference = image.convert("RGB")
with Image.open(image_root / row["target_image"]) as image:
    target = image.convert("RGB")

print(reference.size, target.size)

The official Hugging Face loading guide and Hub download guide provide additional loading and revision-pinning options.

Evaluation protocol

• Metric: top-1 MCQ accuracy for each of Q1–Q7. Avg. is the unweighted arithmetic mean of the seven task accuracies, not a pooled average over all QA rows.
• Cascading scoring: an incorrect foundational answer causes subsequent dependent answers for the same sample to be counted as incorrect. Later-task scores therefore reflect both the task itself and upstream judgments.
• Setting: zero-shot prompting without task-specific fine-tuning on CDBench. Each evaluation is run twice, and the reported scores average the two runs.
• Inputs: distinguish the single-image tasks from paired-image tasks. Keep answers and annotation-only information out of model prompts; report any additional masks, retrieved context, or expert guidance used by a system.

For comparable results, report the evaluated subsets, model version, prompts, input context, per-task scores, and scoring protocol. Independent per-question accuracy without cascading is a different metric. The current repository provides data files; it does not contain a released scoring script or machine-readable task-dependency specification.

Reported model performance

Accuracy (%). Bold marks the best value in each column. These are the reported benchmark runs, not evaluations of the latest versions of the named model families.

Model	Q1	Q2	Q3	Q4	Q5	Q6	Q7	Avg.
Qwen-Max-Latest	57.03	79.97	55.88	72.33	61.78	64.35	65.00	65.19
Qwen-plus-latest	55.12	79.04	55.87	72.18	62.77	63.70	65.93	64.94
Qwen-turbo-1101	48.90	68.15	55.93	66.07	54.63	57.25	59.82	58.68
Grok-3	73.16	80.38	59.67	68.51	60.58	61.70	62.26	66.61
Claude-3-5-sonnet	70.17	82.27	67.47	71.74	64.96	58.48	64.15	68.46
Claude-3-7-sonnet	65.87	75.26	79.87	65.40	61.58	59.92	61.26	67.02
Gemini-1.5-flash	48.21	70.15	57.30	63.51	60.00	55.81	56.15	58.73
Gemini-1.5-pro	72.54	83.07	66.03	73.96	66.55	67.59	68.60	71.19
GPT-4o	91.37	85.00	69.72	69.70	56.83	61.71	60.46	70.68
ChangeAgent	96.87	76.78	78.81	76.79	77.67	70.82	69.99	78.25

ChangeAgent improves the average by 7.06 percentage points over Gemini 1.5 Pro and 7.57 points over GPT-4o. It ranks first on Q1 and Q4–Q7; GPT-4o leads Q2, and Claude-3-7-sonnet leads Q3.

Task-group means, overall means, and individual Q1–Q7 accuracies from the table above. Shading identifies ChangeAgent.

ChangeAgent

ChangeAgent combines paired CLIP visual features, an aggregated change map, retrieved domain knowledge, and task inputs for MLLM reasoning. The change map supplies spatial evidence, while the original visual features retain broader scene semantics. Retrieval adds relevant priors, task definitions, and change examples.

ChangeAgent uses both visual guidance and retrieved knowledge for seven-task reasoning.

Because it uses additional visual guidance and retrieval, ChangeAgent serves as a hybrid reference, rather than an input-matched MLLM baseline. The reported comparison does not isolate the contribution of each component; that requires ablation experiments.

Visual dependence and language bias

GPT-4o was evaluated with and without images on the same randomly sampled 1,000 image pairs (4,628 QA items) under the cascading protocol.

Input	Avg. (Q1–Q7)	Q4	Q5	Q6	Q7
With images	70.68	69.70	56.83	61.71	60.46
Text-only	32.93	26.39	15.74	22.74	21.88

Removing images reduces the average by 37.75 percentage points, showing a substantial contribution from visual evidence. This experiment does not establish the absence of residual language bias. Cascading also propagates upstream errors, so later-task scores should be interpreted in that context.

Scope and limitations

CDBench measures structured visual comparison and answer selection across the included domains. It complements pixel-level change-detection benchmarks and free-form captioning evaluations; its accuracy should not be interpreted as segmentation quality or unrestricted generation quality. Retrieved normal references in anomaly datasets are not necessarily temporal observations of the same object, and the image pairs are not universally registered.

The source datasets, task frequencies, and answer formats are heterogeneous. Report domain and task coverage when comparing systems. Cross-model refinement and expert verification aim to improve annotation quality, but do not guarantee that all ambiguity, language priors, or model-specific bias has been eliminated.

Project and demo

• Project website
• Interactive demo — availability and response times depend on the demo server.
• Community discussions for data issues, missing files, and questions.

License and acknowledgements

This repository is distributed with the CC BY-NC-SA 4.0 license tag. Images originate from the seven source datasets listed above; consult their original licenses and attribution requirements as well.

We thank the creators of LEVIR-CD, SYSU-CD, CDD, MVTec-AD, MVTec-LOCO, VisA, and GoodsAD; the experts who annotated and verified the QA items; Fudan University; Shanghai Ocean University; Shanghai Vision Medical Technology Co., Ltd.; and the open-source community supporting the underlying vision-language models and tools.

Citation

To cite this dataset release:

@misc{chen2026cdbench,
  title  = {{CDBench}: A Comprehensive Multimodal Dataset and Evaluation Benchmark for General Change Detection},
  author = {Chen, Jintao and Qian, Rui and Peng, Bo and Cheng, Linjie and Li, Xiang and Chen, Tian and Ye, Jingyong and Yang, Yinhao and Chi, Mingmin and Dou, Dejing},
  year   = {2026},
  url    = {https://huggingface.co/datasets/MM-CD/CDBench},
  note   = {Dataset and benchmark}
}