README.md

metadata

pretty_name: HVSBench
language:
  - en
task_categories:
  - visual-question-answering
  - image-classification
tags:
  - multimodal
  - vision-language
  - human-visual-system
  - visual-saliency
  - scanpath
  - benchmark
license: other
size_categories:
  - 1K<n<10K
dataset_info:
  features:
    - name: index
      dtype: int64
    - name: image
      dtype: image
    - name: question
      dtype: string
    - name: question_type
      dtype: string
    - name: answer
      dtype: string
    - name: answer_type
      dtype: string
    - name: category
      dtype: string
    - name: split
      dtype: string
    - name: real_image_path
      dtype: string
    - name: origin_dataset
      dtype: string
  splits:
    - name: test
      num_bytes: 2041995103
      num_examples: 8657
  download_size: 1991106611
  dataset_size: 2041995103
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*

HVSBench

HVSBench is a benchmark for evaluating how well multimodal large language models align with human perceptual behavior. It covers human visual system tasks across prominence, subitizing, prioritizing, free-viewing, and searching.

This Hugging Face release packages the divided 10% test subset described in the paper as parquet shards with embedded image bytes. It contains 8,657 question-answer examples and 7,507 unique raw images referenced through the image column. The corresponding paper describes the full HVSBench benchmark with 85,147 multimodal QA pairs across 13 question types and 5 fields.

Paper: Do MLLMs Exhibit Human-like Perceptual Behaviors? HVSBench: A Benchmark for MLLM Alignment with Human Perceptual Behavior

Authors: Jiaying Lin, Shuquan Ye, Dan Xu, Wanli Ouyang, Rynson W. H. Lau

Project page: https://jiaying.link/HVSBench/

Dataset Structure

The dataset is uploaded as parquet shards. The image feature contains the image data embedded in parquet, and real_image_path preserves the original relative source path under raw_datasets.

Columns:

• image: raw RGB image embedded in parquet
• index: source row index
• question: benchmark prompt
• question_type: question type ID, Q1-Q13
• answer: ground-truth answer
• answer_type: answer format, such as single_choice, int_number, sorting, or fixation_prediction
• category: benchmark category, such as saliency_rank, salient_instance, or scanpath
• split: original split label
• real_image_path: original relative path under raw_datasets
• origin_dataset: source dataset name

Data Statistics

• Rows: 8,657
• Unique images: 7,507
• Categories:
  • saliency_rank: 4,431
  • scanpath: 3,162
  • salient_instance: 1,064
• Original source datasets:
  • SIFR: 4,431
  • COCOFreeView: 2,519
  • SIS10K: 1,064
  • COCOSearch: 643

Usage

from datasets import load_dataset

ds = load_dataset("<repo_id>")
sample = ds["train"][0]
image = sample["image"]
question = sample["question"]
answer = sample["answer"]

Citation

@InProceedings{Lin_HVSBench,
  author    = {Lin, Jiaying and Ye, Shuquan and Xu, Dan and Ouyang, Wanli and Lau, Rynson W.H.},
  title     = {Do MLLMs Exhibit Human-like Perceptual Behaviors? HVSBench: A Benchmark for MLLM Alignment with Human Perceptual Behavior},
  booktitle = {CVPR 2026 Findings},
  year      = {2026},
}

License

The dataset is derived from multiple source datasets (SIFR, COCOFreeView, SIS10K, and COCOSearch). Please review and comply with the license and usage terms of the original source datasets when using this release.