Visual Stimuli: Shared1000

  • Source: Natural Scenes Dataset (NSD) shared images (see NSD Experiments documentation)
  • Count: 1,000 images (in images/ subdirectory)
  • Format: PNG
  • Naming: shared{NNNN}_nsd{NNNNN}.pngshared index (1-based, zero-padded to 4 digits) + original NSD image ID
  • Selection: These are the 1,000 images designated as “shared” in the NSD experiment (shown to all NSD participants). Selection criteria and image properties are documented by the NSD group.
  • Licensing: TBD — follows NSD data sharing terms

Image Metadata

nsd_stim_info.csv contains metadata for the full NSD stimulus set (73,000 images). Rows where shared1000=True correspond to the images used in this study.

Column Description
(index) 0-based row index
cocoId MS-COCO image ID
cocoSplit MS-COCO dataset split (e.g., val2017)
cropBox Crop coordinates used to extract the stimulus from the original image
loss NSD loss metric for the image
nsdId NSD image ID — matches the _nsd{NNNNN} portion of the PNG filename
flagged Whether the image was flagged in NSD
BOLD5000 Whether the image appears in the BOLD5000 dataset
shared1000 True for images in this study’s stimulus set

COCO Annotations

All 1,000 shared images originate from the MS-COCO train2017 split. COCO annotations (captions and object instances) have been extracted for these images and stored in two CSV files.

coco_annotations.csv — one row per image (1,000 rows):

Column Description
nsdId NSD image ID (links to filenames and nsd_stim_info.csv)
cocoId MS-COCO image ID
caption_1 through caption_5 Five human-written captions from COCO
object_categories Semicolon-separated list of COCO object categories detected in the image
object_counts Category:count pairs (e.g., person:3; dog:1)
supercategories Semicolon-separated COCO supercategories (e.g., animal; person)
n_object_instances Total number of annotated object instances
n_unique_categories Number of distinct object categories

coco_captions.csv — one row per caption (5,000+ rows):

Column Description
nsdId NSD image ID
cocoId MS-COCO image ID
caption_index Caption number (1–5)
caption The caption text

These files were generated by scripts/extract_coco_metadata.py from the official COCO annotations_trainval2017 release. The 80 COCO object categories span supercategories including person, vehicle, outdoor, animal, accessory, sports, kitchen, food, furniture, electronic, appliance, and indoor.

Computational Features (viz2psy)

All 1,000 shared images are processed with viz2psy, producing a single consolidated output:

  • viz2psy_scores.csv — one row per image (~2,900 columns), indexed by filename
  • viz2psy_scores.meta.json — feature definitions, model versions, and provenance

Features extracted per image include:

Model Columns Description
resmem 1 Image memorability score (0–1)
emonet 20 Emotion category probabilities (e.g., Awe, Joy, Fear)
clip 512 CLIP vision-language embeddings (L2-normalized)
dinov2 768 DINOv2 self-supervised visual features
gist 512 Gabor-based spatial envelope descriptors
places 467 Scene category probabilities (365) + SUN attributes (102)
llstat 17 Low-level statistics (luminance, color, edges, spatial frequency)
caption 1 Natural language image description (BLIP)
saliency 576 Predicted fixation density on a 24x24 spatial grid
aesthetics 1 Aesthetic quality rating (1–10)
yolo 85 Object detection counts (80 COCO classes) + summary stats

See the viz2psy documentation for full column definitions, or consult the .meta.json sidecar.