AOI Analysis Results for Basketry Set: Human vs. Vision Transformer Attention

Areas of Interest analysis results comparing human visual attention with Vision Transformer attention mechanisms for 20 artisanal objects. The dataset contains 1,152,000 distance evaluations from four metrics, derived from eye-tracking data of 30 participants and a pre-trained DINO ViT model. Author Miguel Carrasco uploaded the 5.5 KB XLS file to figshare in April 2026.

Use Cases

• Benchmarking Vision Transformer attention heads against human gaze patterns based on correlation metrics like KL divergence and SSIM.
• Analyzing AI attention concentration within object regions versus background based on AOI lift values.
• Investigating differences in attention strategies between human focal perception and AI global patterns for handcrafted objects.

Strengths

• Data is derived from a controlled experiment with 30 participants and 20 objects, providing a defined comparison basis.
• Analysis includes 1,152,000 evaluations from four complementary statistical metrics (KL divergence, SSIM, CC, SIM).
• Specific attention heads (e.g., #12) are identified as showing strong alignment (+30 to +40 percentage point lift) with human attention.

Limitations

• Row count and column-level documentation are unknown, requiring manual inspection after download.
• The dataset is very small (5.5 KB), indicating limited scope, likely containing summary statistics rather than raw data.
• The sample is limited to 20 specific artisanal objects (basketry bags and ginger jars), which may affect generalizability.

Provenance

Source

Miguel Carrasco via figshare.

Collection Method

Generated from an eye-tracking experiment using a Pupil Labs eye-tracker and analysis of the same objects with a pre-trained Vision Transformer (DINO) model.

Time Range

The study period is not specified; the dataset was published in 2026.

Freshness

Last updated 2026-04-03 17:52:55.

Geography

Geographic coverage of the study or participants is not specified.