AOI Analysis Results for Jar Set: Human and ViT Attention Correlation

A 2026 study by Miguel Carrasco compares human and Vision Transformer (ViT) attention patterns for 20 artisanal objects. The dataset contains results from 1,152,000 distance evaluations across four metrics, based on eye-tracking data from 30 participants and attention maps from a 12-head ViT model. It is a 5.5 KB XLS file shared under a CC-BY-4.0 license.

Use Cases

• Benchmarking Vision Transformer attention heads against human gaze patterns based on the four comparison metrics (KL divergence, SSIM, CC, SIM).
• Analyzing the correlation between AI and human attention for specific object features like buckles in basketry items.
• Investigating differences in global versus focal attention strategies between AI models and human perception.
• Applying findings to product design and aesthetic evaluation tools based on the identified attention head alignments.

Strengths

• Results are based on a controlled experiment with 30 participants and a pre-trained ViT model.
• Analysis includes 1,152,000 distance evaluations across four complementary metrics for robustness.
• Specific attention heads (e.g., #12) are identified as showing strong alignment with human patterns.

Limitations

• Row count and column-level documentation are unknown; field semantics must be inferred after download.
• The dataset is very small (5.5 KB), indicating limited scope and likely summary statistics rather than raw data.
• Data may reflect bias inherent to the specific 20 artisanal objects and participant sample used in the study.

Provenance

Source

Miguel Carrasco via figshare.

Collection Method

Generated from an eye-tracking experiment and analysis of a pre-trained Vision Transformer (ViT) model with DINO.

Time Range

Study conducted circa 2026.

Freshness

Last updated 2026-04-03 17:52:56.

Geography

Geographic coverage of the study is unknown.