SideViewCows2026 – Dairy Cow Re-Identification Dataset

An image dataset for individual re-identification of dairy cows, comprising 80,260 images with corresponding segmentation masks of 110 individual animals, recorded in three different settings. All animals are photographed from the right-hand side. The herd is predominantly Holstein Friesian with a smaller number of Fleckvieh (Simmental), one Jersey, and one Brown Swiss animal. Coat patterns are therefore mostly black-and-white. The dataset is not balanced across breeds. Click here to view example images from all three subsets, one individual per row, with the segmentation mask overlaid. Contents Subset Images Individuals Description parlor 54,393 110 Video frames from a fixed camera at the milking parlor entrance. Consistent viewpoint and framing. barn 25,260 69 Frames from handheld video recorded in the barn. Varying viewpoint and camera motion. snapshots 607 63 Individual photographs in varied settings. Camera angle varies considerably more than in the other subsets; locations differ (indoors and outdoors), and some animals are lying down in their cubicles. The subsets are nested: the 63 individuals in snapshots are a subset of the 69 in barn, which are in turn a subset of the 110 in parlor. Identifiers are consistent across all three subsets, so the same animal carries the same individual_id everywhere. This structure supports cross-domain evaluation. parlor provides a large gallery under controlled conditions, while barn and snapshots offer query images under progressively greater domain shift. The 63 shared individuals are the ones for which images exist in every setting. Directory layout .├── manifest.csv├── SHA256SUMS├── README.md├── preview.jpg├── parlor/│ ├── images/│ │ ├── 128/│ │ │ ├── 128_0001.jpg│ │ │ └── ...│ │ └── ...│ └── masks/│ ├── 128/│ │ ├── 128_0001.png│ │ └── ...│ └── ...├── barn/│ ├── images/│ └── masks/└── snapshots/ ├── images/ └── masks/ Each subset contains images/ and masks/, subdivided into one directory per individual. Files follow the pattern <individual_id>_<nnnn>, numbered consecutively per individual and ordered chronologically. Image and mask share the same filename stem, so the mask for parlor/images/128/128_0001.jpg is parlor/masks/128/128_0001.png. Images are JPEG. Pixels are at native recording resolution — no image has been up- or downscaled — but the parlor frames are cropped to a region of interest rather than showing the full camera frame. Image dimensions therefore vary; the exact size of every image is given in manifest.csv. Masks are PNG, binary (white foreground on black background), the same dimensions as the corresponding image. The foreground is the silhouette of the animal the image belongs to. manifest.csv One row per image. Column Description image_path Path to the image, relative to the dataset root mask_path Path to the corresponding mask individual_id Identifier of the animal, consistent across all subsets subset parlor, barn or snapshots frame_no Consecutive number of the image within that individual and subset time_offset_s Time of capture in seconds, relative to the earliest image in the dataset (see below) width, height Image dimensions in pixels sha256 SHA-256 checksum of the image file time_offset_s Absolute capture dates have been removed. time_offset_s gives the elapsed time in seconds relative to a single global reference point: the earliest parlor frame, which is defined as 0. All intervals between recordings are therefore preserved — both within and between subsets — while no absolute date can be derived. For barn, the value is the start time of the source video and so is identical for all frames of a recording. Privacy measures Regions in which people were detected are overwritten with solid magenta, RGB (255, 0, 255). Detection was carried out with a deliberately recall-oriented procedure and reviewed manually. Owing to JPEG compression, pixels at the edges of these regions may deviate slightly from the exact color value; a tolerance-based comparison locates them reliably. individual_id is a pseudonym and does not correspond to the herd numbers used on the farm. Filenames and directory names contain no capture dates or holding identifiers. Verifying integrity SHA256SUMS covers every file in the dataset: Bash:shasum -c SHA256SUMS Loading the data Python:import pandas as pdfrom pathlib import Pathfrom PIL import Image root = Path(".")df = pd.read_csv(root / "manifest.csv") # All images of one individual, across every subsetsubject = df[df.individual_id == df.individual_id.iloc[0]] # Gallery/query split across domainsgallery = df[df.subset == "parlor"]query = df[df.subset == "snapshots"] row = df.iloc[0]img = Image.open(root / row.image_path)mask = Image.open(root / row.mask_path) Notes on use Frames from the same recording are strongly correlated. When constructing evaluation splits, group by individual and recording rather than sampling frames at random, or performance figures will be optimistic. Contact and maintenance Maintained by Sebastian Möller, Osnabrück University of Applied Sciences — [e-mail], ORCID 0009-0008-5364-6506. Questions about the dataset, corrections, and reports of errors are welcome. If you find an image in which a person is still identifiable, please report it to the address above; such images will be corrected or removed in a revised version. Citation If you use this dataset, please cite it directly: Möller, S. (2026). SideViewCows2026 – Dairy Cow Re-Identification Dataset [Dataset]. Zenodo. https://doi.org/10.5281/zenodo.21605650 The dataset was generated based on the pipeline described in: Möller, S., Hölscher, M., & Morisse, K. (2025). Automatisierte Erzeugung eines Trainingsdatensatzes zur bildbasierten Tieridentifikation mittels KI. In Lecture Notes in Informatics: Proceedings, Vol. 2025, No. 358 (pp. 351–356).Gesellschaft für Informatik e.V. (GI). https://dl.gi.de/handle/20.500.12116/45701