SUM

High-Resolution Fine-Grained Perception Across Controlled Image Views

🤗 Hugging Face · 🤖 ModelScope · English · 中文

700 questions · 29 aligned views · 5 task categories

🔎 Introduction

SUM is an evaluation collection for studying fine-grained visual perception under controlled changes to image context, resolution, and background. Each of its 700 multiple-choice questions is paired with 29 aligned image views, so the same question can be evaluated on the full image, a region crop, a white canvas, a downsampled image, or a view retaining only the target bounding-box pixels.

The collection draws on selected samples from VStarBench (115), ZoomBench (485), and HR-Bench (100). The release contains 20,300 view images, representing 700 questions and 664 distinct original-image file hashes. These are different units: the transformed views are not additional independent questions.

🎯 What can be compared?

View family	Variants	Images	What changes
Original	Full image	700	Reference scene and resolution
Crop	1, 1.5, 3, 6, 12, 18 × bbox area	4,200	Amount of context around the target
Whiteground	The same 6 crop settings	4,200	Crop pasted at its original position on a full-size white canvas
Downsample	11 linear scales from 0.9375 to 0.0625	7,700	Full-scene spatial resolution
Target-only	1, 3, 6, 18 × crop canvas, plus original canvas	3,500	Only target-bbox pixels retained; all other pixels replaced with white

Crop multipliers describe area, while downsample factors describe width and height. These quantities should not be compared as if they were the same scale. The target-only view preserves pixels inside a bounding box; it is not an instance-segmentation mask. Whiteground preserves all content inside the selected crop, including its surrounding context.

Thumbnails are resized for display. Pixel dimensions below each thumbnail describe the stored image. Released benchmark images retain their original encoding and resolution.

📊 Dataset composition

Task	Questions
Color Attributes	347
OCR	118
Structural Attributes	94
Object Identification	84
Material Attributes	57
Total	700

There are 679 four-option and 21 two-option questions. Questions are primarily in English; some OCR targets and a small number of source options contain Chinese or Japanese text. VStarBench samples come from direct_attributes; HR-Bench samples come from the local FSP_8k_100 selection. The category mapping used by this release is preserved in the legacy metadata, including the historical merge of Visual Prompting and Fine-Grained Counting into Object Identification.

📦 Files and configurations

Native Parquet

The default loader reads Apache Parquet: 29 configurations × 700 rows, packaged in 187 shards (44.10 GB). Each row contains the question, answer, view metadata and an image column recognized by Hugging Face Datasets. Image bytes retain their original encoding and resolution. No separate image download is needed when loading a Parquet configuration.

Shards target 256 MiB, and every row group is below 100 MiB. All annotations and available image SHA-256 hashes were checked after conversion. The Parquet manifest records the schema, row counts, shard sizes and checksums.

README.md / README_ZH.md          Dataset cards
parquet/<view>/test-*.parquet     Native Parquet shards (default loader)
metadata/parquet_manifest.json   Parquet schema, row counts and checksums
data/<view>/test/metadata.jsonl   700 image-linked rows per view
data/<view>/test/images/          Original encoded image files
dataset.json                     700 canonical questions; original-view image paths
metadata/image_manifest.json     Per-image SHA-256, byte size and dimensions
metadata/validation_summary.json Full release validation summary
metadata/errata.json              Known issues and historical corrections
metadata/legacy/                  Legacy metadata with portable paths
metadata/path_mapping.json        Mapping from the previous directory layout
scripts/prompts.py                Normalized and legacy prompt constructors
scripts/evaluate.py               Exact option-letter accuracy
assets/                          Logos and data-derived figures
THIRD_PARTY_NOTICES.md            Source attribution and terms

The default configuration is original. Every configuration has a single test split with 700 rows. Examples include crop_3x, whiteground_3x, downsample_0_25, and target_only_original. The complete configuration list appears in the dataset selector and metadata/validation_summary.json.

Field	Meaning
id, base_sample_id	Stable question ID shared across views
image	Image decoded from the Parquet image bytes
image_path, image_sha256	Repository-relative source image path and expected SHA-256
image_available	Whether this row contains image bytes
image_unavailable_reason	Null for available images; otherwise the recorded availability reason
file_name	Image path relative to the corresponding metadata file, in raw JSONL
image_group_id	SHA-256 of the original image file; identical-image questions share this ID
source_dataset, source_question_id, source_image_id	Source provenance
question, question_raw	Normalized question stem and original source question text
choices, answer, answer_text	Option list, correct letter and corresponding text
task_type	One of the five released task categories
view_id, view_family	Exact transformation setting and family
width, height	Dimensions of the current view
bbox_original, bbox_clipped, bbox_view	Source, in-bounds original-image, and current-view coordinates
crop_box_original	Crop or paste region in original-image coordinates
crop_area_multiplier, downsample_linear_scale	Separate parameters for the two scale definitions

Coordinates are pixel-space xyxy. bbox_view for downsampling uses the actual output width and height, including rounding. Legacy metadata retains the crop boxes, effective area multipliers and pixel rules used to generate the released images. Paths beginning with upstream:// inside legacy metadata are provenance identifiers, not additional bundled files.

🚀 Quick start

Hugging Face: load one view

from datasets import load_dataset

dataset = load_dataset("RL-MIND/SUM", "original", split="test", streaming=True)
sample = next(iter(dataset))
print(sample["question"], sample["choices"], sample["answer"])
image = sample["image"]

# Use the same question IDs for a paired comparison.
crops = load_dataset("RL-MIND/SUM", "crop_3x", split="test", streaming=True)

Download an individual sample

import json
from huggingface_hub import hf_hub_download

annotations = hf_hub_download("RL-MIND/SUM", "dataset.json", repo_type="dataset")
with open(annotations, encoding="utf-8") as f:
    sample = json.load(f)[0]
image_path = hf_hub_download("RL-MIND/SUM", sample["image"], repo_type="dataset")

ModelScope mirror

from pathlib import Path
from datasets import load_dataset
from modelscope_hub import HubApi

root = HubApi().download_repo(
    "KAIWANG/SUM", "dataset",
    allow_patterns=["parquet/original/*.parquet"],
)
dataset = load_dataset(
    "parquet",
    data_files={"test": str(Path(root) / "parquet/original/*.parquet")},
    split="test",
)
sample = dataset[0]
image = sample["image"]

Install datasets, huggingface_hub, Pillow, or modelscope-hub as needed. The Parquet version totals approximately 44.10 GB. The original image folders and JSONL annotations remain available (approximately 44.4 GB); choose the format and views you need to avoid downloading both copies.

✅ Evaluation and release notes

• Evaluate the same IDs under each view and report accuracy by source, task and view. scripts/evaluate.py scores explicit option letters; absent or invalid answers count as incorrect.
• The 485 ZoomBench source question strings already include choices. question separates their stem from the choices; question_raw preserves the old text. scripts/prompts.py offers both the normalized prompt and the legacy prompt that repeats those choices. Prompt changes can affect results, so record which version was used.
• Thirty source bounding boxes extend slightly beyond image boundaries. The source coordinates are retained alongside clipped coordinates; source image bytes and existing transformed views are preserved.
• Sample 000669 contains the duplicate distractors red and Red; its recorded answer is C / Brown. The source options are retained and the issue is flagged in quality_notes and the errata. No replacement distractor was invented.
• Repeated original images can support different questions. Keep the same image_group_id and all of its views together if creating any downstream split. Do not treat the 29 paired views as independent data points.
• Every released image was fully decoded, checked against its expected dimensions and SHA-256 hashed during packaging. This structural validation is not a new human review of every semantic answer or a rerun of all transformation pixel-equivalence checks.

📜 Sources, terms and citation

Please credit V*Bench, ZoomBench, and HR-Bench, as applicable. SUM provides a selected, normalized and transformed evaluation collection; it does not claim that the source questions or imagery were newly authored for this release.

Source-specific terms and unresolved upstream license details are documented in THIRD_PARTY_NOTICES.md. The metadata deliberately uses license: other; this is not a blanket permissive license for all underlying images.