Human and Vision Transformer Attention Comparisons for Handcrafted Objects

30 participants viewed 20 artisanal objects while their gaze patterns were recorded with a Pupil Labs eye-tracker. Miguel Carrasco published this dataset in 2026, which compares these human visual attention heatmaps to attention maps extracted from a pre-trained Vision Transformer model using four metrics across 1,152,000 distance evaluations. The analysis identifies specific attention heads, such as head #12, that show the strongest alignment with human visual patterns.

Use Cases

• Benchmarking Vision Transformer attention mechanisms against human eye-tracking data based on the described correlation metrics.
• Analyzing attention head performance for specific object features like buckles based on the Areas of Interest (AOI) analysis mentioned.
• Investigating differences between global AI attention and focal human attention patterns described in the study.

Strengths

• Data is derived from a controlled experiment with 30 participants and a mean age of 24.6 years.
• Analysis includes 1,152,000 distance evaluations using four complementary metrics (KL divergence, SSIM, CC, SIM).
• Specific attention heads, such as #12, are identified with lift values of +30 to +40 percentage points in AOI analysis.

Limitations

• Row count is unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset is very small at 5.5 KB, indicating limited scope.

Provenance

Source

Miguel Carrasco via figshare

Collection Method

Eye-tracking experiment with 30 participants viewing 20 artisanal objects, combined with analysis of a pre-trained Vision Transformer (ViT) model.

Time Range

null

Freshness

Last updated 2026-04-03 17:52:52; freshness should be verified.

Geography

null