Scene8

The original Spatial Envelope collection contains 2,688 RGB outdoor scene images in eight classes, each 256 x 256 pixels. The builder preserves original JPEG bytes, labels, and image IDs. It exposes the complete collection as train; this is not an official training partition.

Eight original Scene8 images, one per outdoor scene class, with class names and image IDs.

Fig. 5 One example per class, selected by the first image ID in each class. Source images are resized for display without cropping or changing labels.

Data Structure

Key

Type

Description

image

PIL.Image.Image

RGB PIL image, 256 x 256 pixels

label

int

Integer class ID (0–7)

image_id

str

Original filename stem, unique within the dataset

Usage

from stable_datasets.images import Scene8

ds = Scene8(split="train")
sample = ds[0]  # image: PIL image, label: int, image_id: str
print(ds.features["label"].names)

# Standard library options control cache locations.
ds = Scene8(split="train", download_dir="/data/downloads",
            processed_cache_dir="/data/processed")

Class IDs follow the fixed order coast, forest, highway, insidecity, mountain, opencountry, street, and tallbuilding. Rows are sorted by image ID. Standard caching, formatting, and transforms are supported.

Testing

The tests download the original release through the standard dataset API and check split sizes, class counts, image decoding, stable IDs, and cache reopening. Offline parser tests additionally check invalid inputs and label matching.

Source and terms

Original page and download: https://people.csail.mit.edu/torralba/code/spatialenvelope/

The original page does not specify a dataset-wide license. The library’s code license does not grant a license to redistribute the images. Download checksums identify the locally audited release rather than publisher signatures.

Citation

@article{oliva2001modeling,
    title={Modeling the shape of the scene: A holistic representation of the spatial envelope},
    author={Oliva, Aude and Torralba, Antonio},
    journal={International Journal of Computer Vision},
    volume={42}, pages={145--175}, year={2001}
}

Known exact-pixel duplicates are retained as source records. Evaluation code should group duplicate images within partitions and avoid train/test leakage.