HueManity: A Benchmark for Testing Human-Like Visual Perception in MLLMs

Paper | Code

Dataset Description

HueManity is a benchmark dataset featuring 83,850 images designed to test the fine-grained visual perception of Multimodal Large Language Models (MLLMs). Each image presents a two-character alphanumeric string embedded within Ishihara-style dot patterns, challenging models to perform precise pattern recognition in visually cluttered environments.

The dataset was created to highlight and address a critical gap: while MLLMs excel at high-level reasoning, their ability to discern subtle visual details often lags significantly behind human performance. HueManity provides a robust framework to measure this nuanced perceptual capability.

Dataset Details

• Total Images: 83,850
• Stimuli: Two-character alphanumeric strings (A-Z, a-z, 0-9). Visually ambiguous characters (e.g., 'I', 'l', 'O') and combinations starting with '0' were excluded from some task sets to prevent evaluation conflicts.
• Image Resolution: $900 \times 900$ pixels.
• Visual Style: Ishihara-style dot patterns, where characters are defined by subtle color differences against a similarly patterned background.
• Color Control: Utilizes 25 meticulously curated foreground-background color pairs. These pairs were selected through a multi-stage process involving CIEDE2000 ($\Delta E_{2000}$) color difference metrics and extensive manual verification to ensure balanced perceptual challenge and human legibility.
• Generation: Images are procedurally generated, and the generation code is provided for reproducibility and further research.

Intended Use & Tasks

HueManity is primarily intended for evaluating and benchmarking the visual perception abilities of MLLMs and other vision models. It can help researchers:

• Quantify model performance on fine-grained pattern recognition.
• Identify architectural or training limitations in current models.
• Drive the development of models with more human-like perceptual robustness.

The paper defines two main evaluation tasks using 1,000-image subsets:

1. Number Recognition Task ("Easier"): Identifying two-digit strings.
2. Alphanumeric Recognition Task ("Harder"): Identifying two-character alphanumeric strings.

Key Findings from Baseline Evaluations

Evaluations presented in the accompanying paper demonstrate:

• Humans: Achieve near-perfect accuracy (100% on numeric, 95.6% on alphanumeric tasks).
• Fine-tuned ResNet50 (Traditional CV): Also performs strongly (96.5% numeric, 94.5% alphanumeric).
• State-of-the-Art MLLMs: Exhibit a significant performance deficit, with the best models scoring around 33.6% on the numeric task and as low as 3% on the alphanumeric task.

Dataset Structure

The dataset consists of:

• Image files (e.g., PNGs).
• Corresponding ground truth labels (the two-character strings).
• Metadata including generation parameters for each image.

Code

You can find the code to generate this dataset at https://github.com/rynaa/huemanity

Citation

If you use the HueManity dataset in your research, please cite the original paper:

@article{grover2025huemanity,
  title={HueManity: Probing Fine-Grained Visual Perception in MLLMs},
  author={Grover, Rynaa and Tamarapalli, Jayant Sravan and Yerramilli, Sahiti and Pande, Nilay},
  journal={arXiv preprint arXiv:2506.03194},
  year={2025}
}