Freiburg Groceries

Overview

FreiburgGroceries loads the authors’ released archive and original train/test lists for 25 grocery categories. Although the paper describes approximately 5,000 images, the downloadable archive contains 4,947 RGB PNG images, each 256 x 256. The file is named .tar.gz but is an uncompressed tar; the loader detects its actual format.

Fig. 3 Five training examples from the Freiburg Groceries release by Jund, Abdo, Eitel, and Burgard (2016). The preview uses the authors’ images and class names. See Source and terms for the original archive, attribution, and research-use conditions.

Official partitions

Select one of five author configurations, split0 through split4. split0 is the default. Both train and test are available; there is no official validation split. Labels and split lists are pinned to author repository commit 8827d2f61c2838756a069562b1c2ff53568ae33b.

Config	Training images	Test images
split0	3,929	1,018
split1	3,989	958
split2	3,962	977
split3	3,937	995
split4	3,943	999

Each published training list has 7,425 rows because the authors oversampled smaller classes for their training procedure. The loader returns each training image path once, in first-occurrence order, preserving the original unbalanced class distribution. Author train/test membership and test ordering are unchanged. Train and test image paths are disjoint within each configuration.

Configurations 2, 3 and 4 omit 8, 15 and 5 archive images, respectively. The loader follows the published membership rather than adding unlisted images. Different configurations are alternative experiments and must not be concatenated as disjoint datasets. Any additional validation split is user-defined.

The release does not provide physical product instance IDs, so path-disjointness does not establish independence of real-world product instances.

Source paths also do not imply unique pixels: the archive contains 14 pairs of different paths with identical decoded RGB arrays (4,933 unique RGB arrays among 4,947 files). Some identical images occur across the published train/test boundary. Applications requiring unseen-image evaluation should audit decoded pixels and explicitly report any experiment-level exclusions; the loader retains the authors’ file membership.

Fields and labels

• image: RGB PIL image by default; float32 CHW tensor with with_format("torch").

• label: integer 0–24, using the authors’ classid.txt order below.

• image_id: relative source path, for example TEA/TEA0123.png.

The label order is not alphabetical:

0 CANDY       1 JUICE       2 VINEGAR      3 OIL         4 CHOCOLATE
5 PASTA       6 RICE        7 MILK         8 SPICES      9 HONEY
10 JAM        11 NUTS       12 CHIPS       13 SODA       14 COFFEE
15 BEANS      16 TEA        17 CORN        18 CEREAL     19 CAKE
20 SUGAR      21 WATER      22 FLOUR       23 TOMATO_SAUCE  24 FISH

Usage

from torch.utils.data import DataLoader
from stable_datasets.images import FreiburgGroceries

data = FreiburgGroceries(
    config_name="split0",
    download_dir="./data/downloads",
    processed_cache_dir="./data/processed",
)
batch = next(iter(DataLoader(data["train"].with_format("torch"), batch_size=8)))
assert batch["image"].shape == (8, 3, 256, 256)

The shared builder caches each configuration separately. Source archives are SHA-256 checked on download. Parsing checks the class vocabulary, path/label agreement, train/test overlap, duplicate test entries, and missing or non-regular image members. Archive paths are never extracted onto the filesystem.

Source and terms

• Author homepage and terms

• Author code and split lists

• Image archive

• Images SHA-256: 35c18d03bb6ed057f12a3c5aea0ac5eef9b8e1d5965d1e32142b7d1f1469e3f0.

• Annotations SHA-256: b07ba6e6c3760cb909f4b92de41f7ab505c3a910f8736613ea4b9712537aafb9.

The authors permit research use, require consent for commercial use, and request citation of their paper. The source image server was accessible over its published HTTP URL; HTTPS was unavailable during verification. A pinned checksum verifies the downloaded bytes. These terms are separate from the library’s software license. The five attributed preview images above are included for dataset documentation; the full source image archive is not included.

Citation

@article{jund2016groceries,
  title={The Freiburg Groceries Dataset},
  author={Jund, Philipp and Abdo, Nichola and Eitel, Andreas and Burgard, Wolfram},
  journal={arXiv preprint arXiv:1611.05799},
  year={2016}
}