CLASSIFYING COW STALL NUMBERS USING YOLO
Dheeraj Vajjarapu
Yeshiva University
New York
dvajjara@mail.yu.edu
Abstract
This paper introduces the CowStallNumbers dataset, a
collection of images extracted from videos focusing on cow
teats, designed to advance the field of cow stall number de-
tection. The dataset comprises 1042 training images and
261 test images, featuring stall numbers ranging from 0 to
60. To enhance the dataset, we performed fine-tuning on
a YOLO model and applied data augmentation techniques,
including random crop, center crop, and random rotation.
The experimental outcomes demonstrate a notable 95.4%
accuracy in recognizing stall numbers.
1. Introduction
Livestock monitoring and management play pivotal roles
in the efficiency and sustainability of modern agriculture.
Among the various aspects of livestock management, accu-
rately tracking and identifying individual animals within a
herd is crucial for optimizing feeding, health monitoring,
and overall farm productivity. In this context, the advent of
computer vision techniques, particularly object detection al-
gorithms, has opened new avenues for automating the iden-
tification and tracking of livestock in real-world agricultural
settings. We have already seen classficiation of coew teats
in [7]
This research focuses on the application of the You Only
Look Once (YOLO) algorithm for the specific task of pre-
dicting cow stall numbers within a barn or farm environ-
ment. The ability to automatically assign stall numbers to
individual cows contributes to streamlined record-keeping,
efficient resource allocation, and targeted health interven-
tions. Leveraging a dataset consisting of 1042 training im-
ages and 261 test images, the model is trained to predict stall
numbers ranging from 0 to 60. YOLO, known for its real-
time object detection capabilities, has shown remarkable
success in various domains, including but not limited to,
pedestrian detection, traffic monitoring, and object recogni-
tion.
The primary objective of this study is to harness the
power of YOLO for accurate and efficient cow stall number
prediction. The reported accuracy of 95% on a test dataset
underscores the potential of the proposed approach to pro-
vide reliable and timely information about the location and
identity of each cow within a monitored space.
As precision agriculture continues to evolve, leveraging
advanced technologies such as YOLO for livestock man-
agement aligns with the broader trend toward automation
and data-driven decision-making. The outcomes of this re-
search not only contribute to the domain of precision live-
stock farming but also have implications for the broader
field of computer vision applications in agriculture.
The subsequent sections of this paper will delve into the
methodology employed, the dataset used for training and
evaluation, experimental results, and discussions on the im-
plications of the findings. Additionally, avenues for fu-
ture research and potential enhancements to the current ap-
proach will be explored. By combining the strengths of
YOLO with the practical challenges and opportunities of
predicting cow stall numbers, this research aims to con-
tribute to the ongoing transformation of agriculture through
innovative and efficient technological solutions.
2. Related Work
Object detection has been a focal point in computer
vision research, with numerous methodologies evolving
over the years. Traditional techniques, such as slid-
ing window-based detectors and region-based CNNs (R-
CNN)[4], marked early attempts at accurate object local-
ization. The introduction of Fast R-CNN [3] and its subse-
quent enhancement with Faster R-CNN addressed compu-
tational inefficiencies, utilizing shared convolutional layers
and a Region Proposal Network (RPN).
Single Shot Multibox Detector (SSD)[5] proposed a
single-shot detection algorithm achieving real-time process-
ing speeds by utilizing multiple feature maps at different
scales. However, the paradigm shift came with the intro-
duction of You Only Look Once (YOLO) by Redmon et
al.. YOLO framed object detection as a regression problem,
dividing the input image into a grid and predicting bound-
arXiv:2401.03340v1 [cs.CV] 23 Nov 2023