Serrelab/SerreOlivaPoggio-PNAS07

Animal / Non-Animal Stimulus Set — Serre, Oliva & Poggio (PNAS 2007)

This dataset is the natural-image stimulus set used in:

Serre, T., Oliva, A., & Poggio, T. (2007). A feedforward architecture accounts for rapid categorization. Proceedings of the National Academy of Sciences, 104(15), 6424–6429. https://doi.org/10.1073/pnas.0700622104

It was the basis for the rapid animal/non-animal categorization psychophysics task in that paper, and has since been re-used in many subsequent studies of rapid visual categorization, feedforward models of the ventral stream, and human–model comparisons.

The original distribution at http://cbcl.mit.edu/software-datasets/serre/SerreOlivaPoggioPNAS07/ is no longer accessible (the CBCL site was reorganized into poggio-lab.mit.edu, a static page that no longer hosts data). This Hugging Face copy is intended as the permanent, citable home for the stimulus set going forward.

Contents

1,200 256×256 RGB JPEG images, organized into two top-level groups × four sub-categories × 150 images:

Targets/        # animal images (600)
  Heads/        # animal head close-ups (150)
  CloseBody/    # whole-body close-ups (150)
  MediumBody/   # animal in scene context (150)
  FarBody/      # small animal, far from camera (150)
Distractors/    # non-animal natural images (600), distance-matched to Targets
  Heads/        # close-up natural objects/textures (150)
  CloseBody/    # close-range natural scenes (150)
  MediumBody/   # medium-range natural scenes (150)
  FarBody/      # far natural scenes (150)

The four sub-categories correspond to the four animal-distance levels described in the paper (head close-up; whole body; animal in scene; small/far animal). Distractor sub-categories were chosen to match the corresponding target sub-category in mean retinal size and scene depth, so that any animal-vs-non-animal effect cannot be attributed to low-level distance/size cues.

The original images were sourced from the Corel Stock Photo Library, the Web, and personal photographs, then cropped and resampled to 256×256 grayscale 8-bit at 7° × 7° of visual angle for the psychophysics task. Files in this release are stored as 3-channel JPEG (the format Reza Ebrahimpour preserved them in); the source content is the same grayscale imagery used in the original experiment.

Loading

from datasets import load_dataset
ds = load_dataset("Serrelab/SerreOlivaPoggio-PNAS07")

Or via direct file access:

from huggingface_hub import snapshot_download
local_dir = snapshot_download(
    repo_id="Serrelab/SerreOlivaPoggio-PNAS07",
    repo_type="dataset",
)

What is not included

• AnimalnessIndex.xls — the original distribution also included a small spreadsheet of per-image "animalness" ratings used in some follow-up analyses. We do not have a copy on hand; if you have one, please open an issue and we'll add it.
• Phase-randomized masks — the rapid-presentation paradigm in the paper used 1/f phase-randomized masks. Those are not part of this image set; they can be regenerated procedurally from the stimuli themselves.
• Per-image metadata — beyond category and sub-category, no additional metadata is shipped here.

Citation

If you use this dataset, please cite the PNAS paper:

@article{serre2007feedforward,
  title   = {A feedforward architecture accounts for rapid categorization},
  author  = {Serre, Thomas and Oliva, Aude and Poggio, Tomaso},
  journal = {Proceedings of the National Academy of Sciences},
  volume  = {104},
  number  = {15},
  pages   = {6424--6429},
  year    = {2007},
  doi     = {10.1073/pnas.0700622104}
}

License

Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).

The source images include material from the Corel Stock Photo Library (a commercial product with restrictive redistribution terms historically), so the dataset is released for non-commercial research and educational use only, with attribution to the PNAS 2007 paper. This mirrors how the stimulus set has been used in the academic community since 2007.

Acknowledgments

Re-supplied by Prof. Reza Ebrahimpour (Sharif University of Technology), who kindly provided his locally-archived copy after the original CBCL link went offline. Re-hosted by the Serre Lab at Brown University.