casperhuggingface/gaaibastu-muzzles

Gaaibastu muzzles

A teaching derivative of the Beef Cattle Muzzle/Noseprint database for individual identification (Yijie Xiong, Guoming Li, Galen Erickson; Zenodo 6324361; DOI 10.5281/zenodo.6324360; CC BY 4.0), used in the Programiz ML Engineering course, chapter 6. The photos show US feedlot beef cattle (Angus, Angus x Hereford and Continental x British crosses), not Nepali cattle. The breed of each animal is not recorded.

Paper: Li, G., Erickson, G. E. and Xiong, Y. (2022). Individual Beef Cattle Identification Using Muzzle Images and Deep Learning Techniques. Animals 12(11), 1453. https://doi.org/10.3390/ani12111453

Contents

Path	What it is
data/muzzles.csv	one row per photo (4,923): index i, cow, file, order_in_cow (file order, the capture-order proxy), role in the course's split, brightness (mean luma), muzzle_luma (median luma of the central box), original width, height
data/dinov2_small.f16.npy	4,923 x 384 float16, facebook/dinov2-small pooler output (layer-normed CLS), row i
data/mobilenetv3_small.f16.npy	4,923 x 1,024 float16, timm/mobilenetv3_small_100.lamb_in1k pooled features
data/dinov2_small_perturbed.f16.npy	14 x 201 x 384 float16, DINOv2-small embeddings of the 201 claim and impostor photos under 14 perturbations
data/dinov2_small_perturbed.json	perturbation names (axis 0) and the photo index i of each row (axis 1)
web/<cow>/<file>	every photo, long side at most 256 px, JPEG quality 85
web/perturbed/	two claim photos shown under each perturbation

Luma is 0.299 R + 0.587 G + 0.114 B on the full-resolution photo. Roles: train and val_claim (214 training cows, oldest 90% / newest 10% of each cow's photos), gallery and claim (43 enrolled test cows, oldest 80% / newest 20%), impostor and unused (11 test cows never enrolled, newest 20% / oldest 80%). Test cows are 20% of the 268 cows by numpy.random.default_rng(42).permutation.

The course's setup: one enrolment photo per cow

The course enrols each insured cow with a single photo, its first in order_in_cow, as an insurer photographs a cow once at policy time. The gallery is every cow outside the 11 never-enrolled ones (257 cows); claims are the newest 20% of the 43 test cows' photos (165); the 11 never-enrolled cows give 36 impostor claims. Top-1 is 76.97% this way against 98.68% on a random split by photo. The role column keeps the earlier split's labels; the one-photo gallery is derived from it as described here.

Chapter 7 files (added 2026-10-16)

Embeddings for chapter 7 of the course (error analysis). All are facebook/dinov2-small pooler outputs (revision ed25f3a3), float16, 384 numbers per image, computed on the original full-resolution photos from Zenodo 6324361 unless a row says otherwise. Row order follows data/muzzles.csv.

Path	What it is
data/dinov2_small_web.f16.npy	4,923 x 384, embeddings of the 256 px web/ copies. One-photo top-1 is 73.94% with these against 76.97% with the originals, so use dinov2_small.f16.npy for course numbers.
data/dinov2_small_brightnorm.f16.npy	4,923 x 384, each photo's RGB scaled so its mean luma is 105.8 (the dataset mean), clipped to 0..255. One-photo top-1 stays at 76.97%.
data/occlusion4.f16.npy	201 x 16 x 384, the 165 claim and 36 impostor photos, each with one square of a 4 by 4 grid set to grey (128)
data/occlusion8.f16.npy	20 x 64 x 384, the same with an 8 by 8 grid, for 10 wrongly matched and 10 correctly matched claim photos
data/occlusion.json	the photo index i of each row (index_4, index_8), the square order (row-major from the top-left), and the share of each square that survives the processor's resize and 224 px centre crop

The grid is drawn on the full photo before the processor resizes the shortest edge to 256 and centre-crops 224. On average only 23% of a 4 by 4 corner square reaches the model, so edge squares show smaller drops partly because they are cut off.

Licence CC BY 4.0, as the source photos.

Licence and changes

CC BY 4.0, as the source. Changes: photos converted to RGB, resized (long side at most 256 px, never enlarged), re-encoded as JPEG quality 85, EXIF removed; embeddings and luma values computed from the original photos; perturbed images in web/perturbed/ are modified for teaching. See ATTRIBUTION.md.