Caltech-101 Silhouettes

Overview

CalTech 101 Silhouettes is a shape classification dataset derived from the Caltech-101 object annotations. Each Caltech-101 image comes with a high-quality polygon outline of its primary object; the outline is centered, scaled and rendered as a filled black polygon on a white background at 28×28 pixels. Outlines producing images that are more than 90% white or black were dropped by the authors.

The dataset uses the authors’ official stratified split (split1). Each class contributes min(3/5 · N_c, 100) images to the training set and the remainder is divided evenly between validation and test, so the evaluation splits are class-imbalanced:

Train: 4,100 images (19-100 per class)
Validation: 2,264 images (6-349 per class)
Test: 2,307 images (6-349 per class)

Because of this imbalance, the authors recommend reporting class-balanced accuracy. A few classes (e.g. soccer ball, pizza, stop sign, yin yang) are indistinguishable by shape alone.

../../_images/caltech101_silhouettes_teaser.png

Data Structure

When accessing an example using ds[i], you will receive a dictionary with the following keys:

Key	Type	Description
`image`	`PIL.Image.Image`	28×28 grayscale binary silhouette (0 = object, 255 = background)
`label`	int	Class label (0-100)

Preprocessing

The source .mat file stores each image as a flattened binary vector in column-major (MATLAB) order with 1-based labels. The builder:

reshapes images in Fortran order so silhouettes are upright,
scales pixel values from {0, 1} to {0, 255}, keeping the original black-object-on-white-background rendering,
shifts labels from 1-101 to 0-100,
keeps class names verbatim from the source classnames field (e.g. "Faces 2", "car side").

Usage Example

Basic Usage

from stable_datasets.images.caltech101_silhouettes import Caltech101Silhouettes

ds_train = Caltech101Silhouettes(split="train")
ds_val = Caltech101Silhouettes(split="validation")
ds_test = Caltech101Silhouettes(split="test")

sample = ds_train[0]
print(sample.keys())  # {"image", "label"}
print(ds_train.info.features["label"].names[sample["label"]])  # "lamp"

# Optional: make it PyTorch-friendly
ds_torch = ds_train.with_format("torch")

References

Homepage: https://people.cs.umass.edu/~marlin/data.shtml
Paper: Inductive Principles for Restricted Boltzmann Machine Learning (AISTATS 2010)
License: No explicit license is stated by the authors. The underlying Caltech-101 annotations are released under CC BY 4.0 (https://data.caltech.edu/records/mzrjq-6wc02).

Citation

@inproceedings{marlin2010inductive,
  title={Inductive Principles for Restricted Boltzmann Machine Learning},
  author={Marlin, Benjamin and Swersky, Kevin and Chen, Bo and de Freitas, Nando},
  booktitle={Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics},
  pages={509--516},
  year={2010},
  volume={9},
  series={Proceedings of Machine Learning Research},
  publisher={PMLR}
}