COIL-20

Overview

../../_images/coil20_teaser.png

COIL-20 contains 1,440 grayscale images of 20 objects. Each object was photographed from 72 views. The processed images are 128 × 128 pixels.

The source dataset does not provide an official train/test split. The train split in this builder contains all 1,440 images.

Data Structure

Each ds[i] returns:

Key

Type

Description

image

PIL.Image.Image

128 × 128 grayscale image

label

int

Object class ID from 0 to 19

view_id

int

View identifier read from the source filename

Usage Example

from stable_datasets.images import COIL20

ds = COIL20(split="train")
print(len(ds))  # 1440

sample = ds[0]
print(sample["image"].size, sample["label"], sample["view_id"])

License and Terms of Use

No dataset-specific license was identified on the official COIL-20 page or in the processed ZIP archive (checked October 29, 2026). The archive contains only image files.

The CAVE general copyright notice states that rights remain with the authors or other copyright holders and that reposting generally requires explicit permission. It does not specify a dataset license for COIL-20. For clarification about permitted uses, contact the COIL maintainers at webcave@lists.cs.columbia.edu, as listed on the official dataset page.

The stable-datasets repository license does not establish a license for the downloaded dataset.

Source and Citation