Open-Set Cattle Muzzle Identification: A Leakage-Controlled Benchmark and Evaluation Protocol

Lalit BC; Dharmendra Singh Chaudhary; Shovit Nepal

cs.CVAug 22, 2026

Open-Set Cattle Muzzle Identification: A Leakage-Controlled Benchmark and Evaluation Protocol

Authors: Lalit BC, Dharmendra Singh Chaudhary, Shovit Nepal

Organizations: Fort Valley State University, School of Environmental, Civil, Agricultural and Mechanical Engineering, Fort Valley, GA, USA · Map Mentors, Department of Electronics and Computer Engineering, Lalitpur, Nepal · Tribhuvan University, Nepal · University of Georgia, USA

Abstract

Reliable individual cattle identification supports disease surveillance, vaccination records, breeding management, and livestock insurance. Although the bovine muzzle provides a stable, non-contact biometric, existing muzzle-recognition systems largely assume a closed set of enrolled animals, limiting their practical deployment. We reformulate cattle muzzle biometrics as an open-set, gallery-based identification problem that can reject previously unseen animals and support incremental enrollment without model retraining. We introduce a leakage-controlled evaluation protocol based on identity-disjoint splits, per-fold retraining, held-out threshold calibration, verified duplicate removal, and bootstrap confidence intervals. We evaluate the framework using two contrasting embedding configurations: a hybrid CNN-ViT metric-learning model and the MegaDescriptor-L foundation model. Under oracle threshold selection, the hybrid model achieves detection-and-identification rates of 98.3%, 96.4%, and 93.6% at target false-acceptance rates of 10^(-1), 10^(-2), and 10^(-3), respectively, while MegaDescriptor-L achieves 99.3%, 98.1%, and 96.1%. However, deployable threshold calibration reveals a substantial difference between oracle and calibrated performance: the hybrid model achieves a false-acceptance rate of 1.03% at a 1% target, whereas MegaDescriptor-L reaches 2.44%. Incremental enrollment further achieves Rank-1 accuracy above 91% with a single reference image and up to 97.3% with eight reference images, without retraining the model or degrading the existing gallery. These results demonstrate that threshold calibration, leakage control, and embedding quality are critical for reliable open-set cattle identification and provide a practical evaluation framework for deployment-oriented animal biometric systems.

Figures & tables

Figure 1: System pipeline overview: (top) muzzle image acquisition and preprocessing, (middle) hybrid CNN-ViT embedding model, (bottom) open-set identification with enrollment and threshold-based rejection.

Stage	Identities	Images
Original database [ 42 ]	268	4923
$-$ fewer than 4 images per identity	0	0
$-$ verified duplicate identities	19	280
Retained benchmark	249	4643

Table 1: Dataset provenance.

Figure 2: Hybrid CNN-ViT embedding architecture: (left) dual-branch feature extraction with cross-attention fusion, (right) gated pooling and quality weighting for robust embedding.

Figure 3: Leakage-controlled protocol with identity-disjoint splits: known (124), validation-unknown (50), and test-unknown (75) identities are partitioned separately to prevent information leakage and ensure valid evaluation at strict false-accept rates.

Metric	Hybrid CNN–ViT	MegaDescriptor-L
Rank-1 (%)	$99.53\;[99.03,99.86]$	$99.86\;[99.72,99.97]$
Verification EER (%)	$0.18$	$0.07$
Open-set EER (%)	$2.77\;[1.90,3.50]$	$1.63\;[1.02,2.06]$
DIR@ $10$ (%)	$98.31\;[97.46,99.01]$	$99.29\;[98.81,99.61]$
DIR@ $10$ (%)	$96.37\;[95.07,97.64]$	$98.10\;[97.26,98.92]$
DIR@ $10$ (%)	$93.59\;[92.18,95.79]$	$96.07\;[95.32,98.10]$

Table 2: Open-set identification, plain cosine, oracle threshold.

Figure 4: Open-set classification rate (OSCR) curve for hybrid CNN-ViT model: plot of true positive rate vs. false positive rate at multiple operating thresholds.

Figure 5: Detection-and-identification rate (DIR) curve for hybrid CNN-ViT model at target false-accept rates of

10

10

, and

10

Figure 6: Open-set classification rate (OSCR) curve for the MegaDescriptor-L foundation model: demonstrates consistent protocol validity across different embedding architectures.

Figure 7: Score distribution comparison: hybrid CNN-ViT model with known and unknown sample scores, illustrating the separation between enrolled and unenrolled animals.

Figure 8: Score distribution comparison: MegaDescriptor-L foundation model showing clean separation between known and unknown score distributions.

Score	$10$	$10$	$10$	OSCR-AUC
Plain cosine	99.29	98.10	96.07	0.9953
Entropy	99.29	98.10	96.04	0.9953
AS-norm	99.29	97.85	94.92	0.9953
Combined	99.21	97.31	93.69	0.9955
Margin	99.21	97.13	92.97	0.9954
OpenMax	91.96	90.32	89.23	0.9354

Table 3: Scoring-function ablation (foundation embedding; DIR@FAR %, five-fold mean).

Figure 9: Threshold transferability across folds: per-fold performance when threshold calibrated on val-unknown is transferred unchanged to test-unknown, showing deployment consistency.

	$n$	Rank-1 new	DIR@ $\tau$ new	Rank-1 orig.	Reject unk.
Hybrid	1	91.08	75.63	99.39	98.83
	3	95.94	86.22	99.39	98.76
	5	96.35	88.28	99.38	98.80
	8	97.07	89.37	99.38	98.84
Foundation	1	92.61	77.54	99.80	98.86
	3	96.46	86.38	99.77	98.86

Table 4: Incremental enrollment of previously unseen identities (no retraining).

n

: enrollment images per new identity.

Figure 10: Incremental enrollment demonstration: gallery-based enrollment with no retraining enables adding new identities while maintaining performance on the existing gallery.

Figure 11: t-SNE visualization of embedding space: hybrid CNN-ViT model showing known identities (colored clusters) and unknown samples (gray), demonstrating the separation necessary for open-set rejection.

Figure 12: Failure modes analysis: representative examples of misidentification and false rejections from the hybrid model, highlighting challenging cases with poor image quality, occlusion, or extreme poses.

Biometric Identification

Biometric identification matches or verifies people and animals using distinctive physical, physiological, or behavioral signals.

See all 128 papers

ReCowGnition: A Realistic Biometric Benchmark for Cow Face Recognition
Jul 24, 2026Marco Huber, Marco Kiesewalter, Judith Louise Pieper +2Biometric Identification Animal Re-Identification
Fraunhofer Institute for Computer Graphics Research IGD, Darmstadt, Germany · Technische Universität Darmstadt, Darmstadt, Germany · Fraunhofer Institute for Computer Graphics Research IGD, Rostock, Germany
Advanced Machine Learning and Deep Learning Techniques for Enhanced Cattle Identification and Detection: A Comprehensive Review
Apr 5, 2026Fayazunnesa Chowdhury, Syed Md. Galib, Md Nasim Adnan +3Agricultural Image Analysis Animal Re-Identification
Jashore University of Science and Technology, Jashore, Bangladesh · University of Information Technology & Sciences (UITS), Dhaka, Bangladesh. · Gazipur Agricultural University, Gazipur, Bangladesh +1
Automated Re-Identification of Holstein-Friesian Cattle in Dense Crowds
Feb 17, 2026Phoenix Yu, Tilo Burghardt, Andrew W Dowsey +1Precision Livestock Farming Animal Re-Identification
School of Computer Science, Merchant Venturers Building, Woodland Road, University of Bristol, Bristol, BS8 1UB, Bristol, United Kingdom · Bristol Veterinary School, University of Bristol, Langford House, Dolberry, Churchill, Bristol, BS40 5DU, Bristol, United Kingdom

Jul 24, 2026J/K move · Enter open · S save

ReCowGnition: A Realistic Biometric Benchmark for Cow Face Recognition

Marco Huber, Marco Kiesewalter, Judith Louise Pieper, Bastian Kubsch+1

Fraunhofer Institute for Computer Graphics Research IGD, Darmstadt, Germany · Technische Universität Darmstadt, Darmstadt, Germany · Fraunhofer Institute for Computer Graphics Research IGD, Rostock, Germany

Biometric Identification Animal Re-Identification Face Recognition

Open paper AskPDF arXiv

With the development of precision livestock farming and the advances in computer vision, visual animal biometrics has gained attention. Using biometric technologies that have been proven effective for humans to identify livestock can increase animal welfare as well as production efficiency. However, challenges such as complex scenarios, similar appearances, occlusions, and non-cooperative behavior, as well as the limited amount of publicly available labeled datasets, remain. In this work, we contribute a novel, publicly available cow face benchmark dataset that has been collected in a realistic automatic scenario with 6,838 images of 161 different cows at a dairy farm. In addition to the public dataset, we define two verification and four identification evaluation protocols to foster comparable research in the cow recognition research field. Further, we provide evaluation results on our dataset of six benchmark models, which include models trained on limited data, cross-species fine-tuned models, and zero-shot foundation model approaches.

Abstract

Biometric Identification

Explore similar work

ReCowGnition: A Realistic Biometric Benchmark for Cow Face Recognition

Advanced Machine Learning and Deep Learning Techniques for Enhanced Cattle Identification and Detection: A Comprehensive Review

Automated Re-Identification of Holstein-Friesian Cattle in Dense Crowds