HARD-VQA

Ultra-High-Resolution Aerial VQA with Original Images Embedded per Question

🤗 Hugging Face · 🟣 ModelScope · 📊 Statistics

English | 中文：Hugging Face · ModelScope

📚 Introduction

HARD-VQA packages ultra-high-resolution aerial visual question answering data as self-contained Parquet shards. Each row is one multiple-choice question, with all required original JPEG bytes embedded in its ordered images column. Single-image, two-image, and three-image questions retain their original image order.

The native package contains 1,563 valid questions across 8 task types, using 854 distinct original images from NJU-xianlin, UAV-BASE-13, and UAV-BASE-14. Every original image is 12,768 × 9,564 pixels. Loading a row requires no external image directory, image URL, or source server: each embedded image has a null path and its complete JPEG bytes.

HARD-VQA and NJU-HARD use the same valid questions and original images, with different storage layouts. HARD-VQA embeds the images required by each question directly in that row; images shared by several questions are stored repeatedly. NJU-HARD on Hugging Face / ModelScope stores each original once in an images configuration and links to it from a separate questions configuration.

📊 Dataset at a Glance

Item	Native package
Valid questions	1,563
Task types	8
Distinct original images	854
Ordered image occurrences	2,571
Questions with 1 / 2 / 3 images	855 / 408 / 300
Original resolution	12,768 × 9,564 pixels
Distinct source JPEG bytes	87,637,348,481 bytes
Embedded JPEG occurrences before Parquet compression	263,224,331,477 bytes
Original-resolution Parquet	391 shards; 263,218,065,108 bytes (approximately 263.22 GB)
Default configuration / split	default / unsplit
Browsing preview	8 examples, one per task, with derived thumbnails

These totals describe the native package. Repeated images account for the larger per-question package size. Original JPEGs are not resized, cropped, or re-encoded; the preview thumbnails are separate browsing aids.

Source task label	Questions
classification	250
counting	200
property	154
spatial_relation	100
task1_specialdetect	66
task3_region_single	151
task4_findit_nearbox	342
task_2_relocate	300
Total	1,563

📦 Data Organization

Configuration	Split	Contents
default (default)	unsplit	Question records with all required original images embedded per row
preview	sample	Eight examples with embedded thumbnails for browsing

data/        # 391 original-resolution Parquet shards
preview/     # vqa_preview.parquet: 8 thumbnail examples
provenance/  # checksums, excluded records, and validation reports
statistics.json

The original-resolution schema is:

Column	Type	Meaning
question_id	int32	Original question ID
images	Sequence(Image()) / List(Image())	Complete original JPEG bytes, ordered as Image 1, Image 2, Image 3
question	string	Original question text
options	struct of A/B/C/D strings	Original choices; C/D are null for binary questions
answer	string	Source-provided correct option key
answer_text	string	Text selected by the source answer key
task_type	string	Original task label
num_images	int32	Number of ordered images
source_image_paths	list of strings	Provenance identifiers; not required to load images
reference_bbox	list of int64	XYXY pixel coordinates, when supplied
scene, sequence	string	Image sequence identifiers
quality_flags	list of strings	Recorded source-quality issues

The Parquet feature metadata uses the compatible Sequence(Image()) representation, which newer Datasets versions normalize to List(Image()).

🚀 Quick Start

Use datasets>=3.6. The package has been checked with Datasets 3.6.0 and 5.0.0. Start with the small preview:

from datasets import load_dataset

preview = load_dataset("RL-MIND/HARD-VQA", "preview", split="sample")
print(preview[0]["question"], preview[0]["options"])

Stream original-resolution examples while keeping the images as JPEG bytes:

from datasets import Image, Sequence, load_dataset

ds = load_dataset(
    "RL-MIND/HARD-VQA", "default", split="unsplit", streaming=True
)
ds = ds.cast_column("images", Sequence(Image(decode=False)))
row = next(iter(ds))
print(row["question"], row["options"], row["answer"])
original_jpeg_bytes = row["images"][0]["bytes"]

Streaming avoids downloading the entire package before iteration. Each original image is large, so inspect one example at a time and allow adequate memory before decoding. To decode a known packaged original:

from io import BytesIO
from PIL import Image as PILImage

PILImage.MAX_IMAGE_PIXELS = None  # Known high-resolution images verified during packaging.
image = PILImage.open(BytesIO(original_jpeg_bytes))

For a local snapshot downloaded from either Hugging Face (RL-MIND/HARD-VQA) or ModelScope (KAIWANG/HARD-VQA), use the same Parquet loader:

ds = load_dataset(
    "parquet",
    data_files={"unsplit": "/path/to/HARD-VQA/data/*.parquet"},
    split="unsplit",
)

A full original-resolution load transfers approximately 263.22 GB and creates a local cache; allow additional cache space.

🔎 Preview and Original Resolution

The preview configuration uses 1,024-pixel thumbnails and marks each row with preview_only=true. It is intended for browsing. Questions retain their original wording; the thumbnails are not the original-resolution evaluation data.

For viewers that do not display image lists, the preview also exposes image_1, image_2, and image_3 as individual image columns. Missing second or third images are null; the ordered images column remains available. Use default for the original images and original-coordinate annotations.

✅ Source Audit and Splits

The source contained 1,565 records. IDs 476 and 496 had null questions, options, and answers because generation parsing failed. They are excluded and recorded in excluded_records.jsonl; images used only by those rejected records are not included.

ID 597 is preserved and flagged duplicate_option_text: options A and B are both white, and the source answer is B. Filter quality-flagged rows when evaluation requires a unique correct option.

Question text, options, answer keys, image order, and supplied bounding boxes are preserved. Structural validation does not establish the semantic correctness of every source answer. The source did not define train/validation/test membership, so the original-resolution package uses one unsplit partition. Shared images and adjacent frames should be considered when creating downstream splits.

🔐 Integrity and Provenance

The packaging process checks embedded image bytes against the source with SHA-256 and records original-image and Parquet checksums. The validation reports document question-field preservation, ordered image references, and schema compatibility:

• Original-image checksums
• Parquet shard checksums
• Validation report
• Schema compatibility report
• Package statistics

The same files are included under provenance/ and at statistics.json in the ModelScope package.

📜 License

The source did not include a license statement, so the dataset card retains license: unknown. This packaging does not grant a new reuse license; contact the dataset owner for usage terms.