Intel Image Classification

Overview

Intel Image Classification is a natural-scene image classification dataset containing photographs in six categories:

  • buildings

  • forest

  • glacier

  • mountain

  • sea

  • street

The dataset was originally published for an Intel image classification challenge.

Intel scene images showing buildings, forest, glacier, mountain, sea, and street classes

Dataset Information

Property

Value

Task

Image classification

Number of classes

6

Original labeled training images

14,034

Original labeled test images

3,000

Builder training images

12,632

Builder validation images

1,402

Total labeled images

17,034

Image format

RGB images

Splits

The builder preserves the original labeled test set.

The original training set is divided into training and validation partitions using deterministic, class-stratified image-path hashing. Within each class, full ZIP-member paths are sorted by (SHA256("42:" + path), path). The first floor(0.1N) members form validation; the rest form training. Split seed 42 is independent of training seed 0. Output paths are sorted, so ZIP entry order does not affect membership. A change to member path prefixes may affect assignments. Test membership is copied directly from seg_test and never included in this partitioning.

The original unlabeled prediction set is excluded because it does not provide class labels for supervised evaluation.

Usage

from stable_datasets.images import IntelImageClassification

ds_train = IntelImageClassification(split="train")
ds_valid = IntelImageClassification(split="validation")
ds_test = IntelImageClassification(split="test")

ds_all = IntelImageClassification(split=None)

print(len(ds_train))
print(len(ds_valid))
print(len(ds_test))

Each example contains:

  • image: The natural-scene image.

  • label: An integer class identifier from 0 to 5.

Source

Class mapping

Labels 0 through 5 mean buildings, forest, glacier, mountain, sea, and street, respectively, for every split.

Download, cache and RGB behavior

The original public Kaggle ZIP is downloaded and read directly without extraction. Original RGB bytes are preserved; any non-RGB images are converted to RGB and encoded as lossless PNG. First loads build processed split caches; repeated loads reopen them offline.

Default directories are ~/.stable-datasets/downloads and ~/.stable-datasets/processed; override them through download_dir and processed_cache_dir, or set STABLE_DATASETS_CACHE_DIR.

Attribution and usage terms

Credit the Intel Image Classification challenge and Puneet Bansal’s Kaggle distribution. The downloaded ZIP has no license file. This integration does not assign a license or transfer the library’s code license to the photographs. Consult the source’s current terms before redistribution; no source photographs are included here.

Citation

@misc{bansal2019intel,
  author={Bansal, Puneet},
  title={Intel Image Classification},
  year={2019},
  howpublished={Kaggle},
  url={https://www.kaggle.com/datasets/puneet6060/intel-image-classification}
}