Susan0803/visual-perception-benchmark-anon

Anonymous Visual Perception Benchmark

This evaluation dataset contains 2,876 image-question pairs with Chinese and English questions, reference answers, and hierarchical visual-perception labels. Images are embedded in Parquet files; no external image service is required.

Data Format

The dataset has one test split. Each row contains:

Field	Description
id	Release-local sample identifier, unrelated to internal identifiers.
image	Embedded PNG image, decoded as a Hugging Face Image feature.
question_zh, question_en	Chinese and English question text.
answer_zh, answer_en	Chinese and English reference answers.
answer_mask	JSON-encoded geometric annotations where provided; otherwise an empty string.
l1_zh through l4_zh	Chinese hierarchical category labels.
l1_en through l4_en	Corresponding English category labels.

The seven top-level categories are Grounding, Semantic Recognition, Fine-grained Discrimination, Visual Tracking, Space Understanding, OCR, and Structured Visual Perception. There are 21 distinct top-level/subcategory pairs. The OCR portion includes 300 questions labeled Text Recognition in Less Common Languages.

The original annotation text is preserved. Some lower-level labels are absent: 7 rows have an empty third-level label and 75 rows have an empty fourth-level label. Empty annotations are not inferred or fabricated.

Loading

from datasets import load_dataset

dataset = load_dataset("Susan0803/visual-perception-benchmark-anon", split="test")
sample = dataset[0]
image = sample["image"]
question = sample["question_en"]
reference_answer = sample["answer_en"]

To load a local checkout, use:

dataset = load_dataset("parquet", data_files={"test": "data/test-*.parquet"}, split="test")

Release Notes

Only questions, reference annotations, images, and category labels are included. Model predictions, model identifiers, evaluation scores, reviewer notes, internal storage addresses, and internal sample identifiers are excluded.

Images are stored as self-contained PNGs without embedded EXIF or text metadata. Image orientation is applied to pixels, and embedded color profiles are converted to sRGB. Multi-picture JPEG containers are represented by their primary image. This removes file metadata, not text visibly present within benchmark images. The images themselves are evaluation content and are not redacted.

This repository does not assert an additional license over third-party image content. Public accessibility alone does not establish redistribution rights.