Metadata Acquisition

Todo

Introductory narrative (2-3 sentences): What metadata was collected or computed for the stimuli beyond basic image properties? Why is this metadata useful (e.g., for controlling confounds, building feature spaces, analyzing category structure)?

Image-level Metadata

Todo

Document all metadata fields that were collected or computed per image. For each field or group of fields, describe:

• What it is and how it was obtained (manual annotation, automated tool, model extraction, database lookup)

• Software / model / API used (with version)

• Units or value range

Candidate metadata categories to cover:

Low-level Visual Properties

Todo

E.g., luminance, contrast, spatial frequency, color statistics. How were these computed? Which tool or script?

Semantic Annotations

Todo

E.g., object categories, scene types, number of objects, animacy. Were these human-annotated, from an existing dataset (COCO, ImageNet labels), or model-derived?

Model-derived Features

Todo

E.g., CLIP embeddings, DNN feature vectors, caption embeddings. Which models and layers? How are they stored (separate files, columns in stimuli.tsv)?

Other Metadata

Todo

Any other metadata: image source/provenance, licensing info per image, NSFW scores, aesthetic scores, etc.

Metadata File Format

Todo

Where does the metadata live? Is it all in stimuli/stimuli.tsv, or are there separate files for different metadata types (e.g., embeddings as .npy)? Document file paths and formats.

Cross-reference Stimulus Data for the full stimulus file organization.