electronics
Article
Study on Body Size Measurement Method of Goat and Cattle
under Different Background Based on Deep Learning
Keqiang Li 1,2,3,4,† and Guifa Teng 5,6, *






Citation: Li, K.; Teng, G. Study on
Body Size Measurement Method of
Goat and Cattle under Different
Background Based on Deep Learning.
Electronics 2022, 11, 993. https://
doi.org/10.3390/electronics11070993
Academic Editor: Amir Mosavi
Received: 7 March 2022
Accepted: 21 March 2022
Published: 23 March 2022
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
Copyright: © 2022 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Electronics 2022, 11, 993. https://doi.org/10.3390/electronics11070993
1 School of Mechanical and Electrical Engineering, Hebei Agricultural University, Baoding 071001, China;
lkq910802@163.com
2 School of Mathematics and Information Technology, Hebei Normal University of Science and Technology,
Qinhuangdao 066004, China
3 Hebei Key Laboratory of Specialty Animal Germplasm Resources Exploration and Innovation,
Hebei Normal University of Science and Technology, Qinhuangdao 066004, China
4 Hebei Agricultural Data Intelligent Perception and Application Technology Innovation Center,
Hebei Normal University of Science and Technology, Qinhuangdao 066004, China
5 School of Information Science and Technology, Hebei Agricultural University, Baoding 071001, China
6 Hebei Key Laboratory of Agricultural Big Data, Baoding 071001, China
* Correspondence: tguifa@126.com
† Affiliations 3 and 4 are under planning.
Abstract: The feasibility of using depth sensors to measure the body size of livestock has been
extensively tested. Most existing methods are only capable of measuring the body size of specific
livestock in a specific background. In this study, we proposed a unique method of livestock body size
measurement using deep learning. By training the data of cattle and goat with same feature points,
different animal sizes can be measured under different backgrounds. First, a novel penalty function
and an autoregressive model were introduced to reconstruct the depth image with super-resolution,
and the effect of distance and illumination on the depth image was reduced. Second, under the
U-Net neural network, the characteristics exhibited by the attention module and the DropBlock were
adopted to improve the robustness of the background and trunk segmentation. Lastly, this study
initially exploited the idea of human joint point location to accurately locate the livestock body feature
points, and the livestock was accurately measured. According to the results, the average accuracy
of this method was 93.59%. The correct key points for detecting the points of withers, shoulder
points, shallowest part of the chest, highest point of the hip bones and ischia tuberosity had the
percentages of 96.7%, 89.3%, 95.6%, 90.5% and 94.5%, respectively. In addition, the mean relative
errors of withers height, hip height, body length and chest depth were only 1.86%, 2.07%, 2.42% and
2.72%, respectively.
Keywords: body size measurement; body segmentation; feature point location
1. Introduction
As precision animal husbandry is leaping forward, scientific feeding and management
can help producers gain insights into the health and overall welfare of livestock. Accord-
ingly, body measurement turns out to be a vital indicator to assess fat deposition and
energy metabolism of livestock [1]. Conventional measurement methods mainly exploit
tools for the measurement. Besides the time-consuming characteristic and high costs, live-
stock will come into contact with measurement processes. The stress response of livestock
affects measurement accuracy while causing personal injury [2]; therefore, it is necessary to
develop an automated measurement method instead of manual measurement.
Against the defects of conventional methods, computer vision has been adopted
in livestock body size measurement [3]. The feasibility of using 2D images has been
proven to measure the body size of goat [4,5], pig [6,7], cattle [8,9], horses [10], and other
https://www.mdpi.com/journal/electronics
Electronics 2022, 11, 993 2 of 22

livestock; however, measurements by 2D images without depth information were often

inaccurate [11,12]. With the emergence of depth sensors (e.g., Kinect, Intel Realsense and
ASUS-Xtion), 3D images with depth information have been used for more accurate body
size measurement of livestock.
For pigs, Wang et al. [13] and Shi et al. [14] used random sample consensus to remove
the background point cloud, and the body measurements were acquired by calculating
the integral length of the curve. For cattle, Salau et al. [12] manually marked the feature
points and calculated the cattle body size with the Cartesian coordinates. The feature point
marking method was then optimized by Ruchay et al. [15]; the filters were used to remove
the noises in the depth map and marked feature points in cattle. For goats, Lina et al. [16]
detected the feature points by finding the convex to calculate the goat’s body size. Guo
et al. [17,18] developed the LSSA_CAU system to measure the body sizes of pigs, horses
and cattle in their respective scenes, exploiting the unevenness of the first derivative of the
point cloud curve to find the feature points of livestock; however, all the above research
on livestock body size measurement analyzed a single livestock species in a specific scene,
and the measurement accuracy was greatly influenced by illumination, background and
livestock species, and therefore it does not have wide application value.
To obtain a widely applicable method based on existing studies, the related research
in livestock body size measurement should focus on how to overcome the influence of
illumination, background and livestock species to obtain high-quality point clouds and
accurate feature points. Numerous researchers have completed background segmentation
and body-part segmentation of livestock by deep learning [5,19–22]; however, this research
has also been limited to single livestock species and has poor robustness to livestock
species and background noise. Further research is needed for multi-species segmentation
by deep learning. For feature point location, feature point could be located by image
segmentation [9,23,24] and statistical learning [14,25]. The accuracy of image segmentation
and statistical learning is determined by feature engineering construction. Compared to
manual feature engineering construction, deep learning is capable of extracting better
feature expressions. Currently, the research on using deep learning to locate the feature
points in livestock is rare.
As revealed from the mentioned research, various livestock exhibit unique morpho-
logical characteristics; however, most methods only apply to the body size measurement of
one species of livestock, and they are affected by background, illuminations and livestock
species. Though deep learning technology has made great progress, it needs further study
in the livestock body size measurement; thus, this study developed a measurement method
of body size for different livestock species based on deep learning. For adult livestock with
the same body size feature points, the deep learning technology was used to complete the
livestock background segmentation, the trunk segmentation and the feature point location.
This study could be referenced for using the deep learning technology to measure the body
size of livestock.
The rest of this study is organized as follows. In the Section 2, the materials and meth-
ods are described. In the Section 3, the experimental results are presented and discussed.
In the Section 4, the conclusion of this study and subsequent studies are summarized.

2. Materials and Methods

2.1. Data Set
2.1.1. Data Acquisition
In this study, used the Intel RealSense D415 depth camera (Intel™, Santa Clara, CA,
USA) with both handheld and fixed shooting to collect video data of 30 adult cattle (Sim-
mental cattle and Holstein Friesian, age > 36 months) and 25 adult goats (Chengde hornless
goat, age > 24 months) in April and November. Animal videos were collected in different
illumination, different distances and different angle under the scope of the camera’s depth
horizontal field of view (±69.4◦ ), color horizontal field of view (±69.4◦ ) and the acquisition
range of the depth camera (16–200 cm). The camera’s field of view was ensured to cover
Electronics 2022, 11, 993 3 of 22

the feature points of the livestock. The average video duration of the respective livestock
was 120 s. The resolution of color video and depth video reached 1280 × 720 pixels, and
the frame rate of video was set to 15 frames/s. To expand the training set of semantic
segmentation model, this study used crawler technology to collect considerable images of
cattle and goats. Figure 1 illustrates the captured image.

Figure 1. Collected image: (A) is the image collected with the depth camera; (B) is the image collected
with the crawler.

At the end, two frames of images were randomly selected from each second of the
video and combined with the network images to construct the final data set. The data set
contains 9895 color images of cows and 9577 color images of goat. The ratio of the training
set, validation set, and test set is 5:3:2.
This experiment does not involve a livestock slaughter experiment, and the data
collection process of related livestock conformed to the regulations of the animal Ethics
Committee of Hebei Agricultural University.

2.1.2. Data Annotation

In this study, two experienced technicians used the LabelImg Annotation Tool (Version
1.8.6) based on PyTorch framework (Version 1.5.0) to manually annotate the characteristic
parts of the color images in the data set. The labels include the torso and feature points
of cattle and goats. The LabelImg annotated the target object in the original image, and
then a JSON file was created for the annotated image. The semantic segmentation model
was used to segment the feature parts of cattle and goats, and the edge contour and feature
points were extracted from the local image of the feature parts.
This section may be divided by subheadings; it should provide a concise and precise
description of the experimental results, their interpretation, as well as the experimental
conclusions that can be drawn.

2.2. Nomentclature of Body Traits and Measurement

During the measurement, the height measuring instrument (the maximum measuring
length, width and height are, respectively: 180 cm, 50 cm and 120 cm, while the instrument
accuracy is one centimeter) was adopted to measure the body size of the cattle and goat,
as shown in Figure 2A. The height measuring instrument was aligned to both sides of the
feature points to obtain the length of the cattle and goat. In order to reduce the manual
measurement error, two surveyors measured each cattle and goat three times. Finally, the
average value was calculated as the final manual measurement result. All length data were
in centimeters and one decimal place was reserved for all measurement results. Withers
Electronics 2022, 11, 993 4 of 22

height, chest depth, hip height and body length should be measured [17,18], as shown in
Figure 2B.

Figure 2. (A) Manual measurement of cattle and goat body dimensions; (B) Illustration of four
measured body dimensions: Line (a,f) is withers height, line; (b,e) is body length, line; (c,h) is chest
depth line; (d,g) is hip height.

Above, a is the highest point of the withers, b represents the anterior extremity of the
hummers, c is the shallowest part of the chest, d expresses the highest point of the hip
bones, and e is the extremity of the ischium. The straight line a–f represents withers height,
the vertical distance from the highest point of the withers to the ground surface at the level
of the forelegs. The straight line c–h expresses chest depth. It is the vertical distance from
the back to the shallowest part of the chest. The straight line d–g refers to hip height. It is
the vertical distance from the highest point of the hip bones to the ground surface at the
level of the hind legs. The straight line b–e denotes the body length; it is the distance from
the anterior extremity of the humerus to the posterior internal extremity of the ischium.

2.3. Image Processing

The ultimate goal of image processing is to get the feature points of livestock in a
color image, and then calculate the body size of livestock combined with a depth image.
The workflow comprised: (1) depth image restoration. Resolution reconstruction of depth
image was performed by combining color image and infrared image. We solved the effect
of illumination and distance on the depth image, and reduced the texture copy artifacts
in the depth image. (2) Set all image sizes to 768 × 768 pixels for data annotation and
network training. The training set, validation set, and test set were divided according
to the proportion of 5:3:2. (3) Body-part segmentation. Based on the optimized U-Net
neural network, the remote, noncontact and automatic background segmentation and trunk
segmentation of cattle and goat were realized, respectively. (4) Feature point location. We
identified and located feature points through the multistage densely stacked hourglass
network. None of the methods mentioned above use any pre-training network. The
different steps are described in the subsections below.

2.3.1. Depth Image Restoration

The depth camera is capable of collecting the depth information of the scene in
real-time; however, illuminations, distances and other environmental factors significantly
impact the quality of depth images [26,27]. To improve the robustness of distance and
illumination and reduce the depth image texture and copy artifacts and depth discontinuity,
this study proposed a depth image super-resolution reconstruction algorithm based on
the autoregressive model. The algorithm consisted of data items and regular items, as
expressed in the following equation. To be specific, the depth map D0 was obtained from
Electronics 2022, 11, 993 5 of 22

the real depth image via the bicubic interpolation. “Ω” denotes the points in the depth
image, and N(i) represents the points beside the point i in the depth image.
  
 2 
D = argminDi ∑ Di − Di0 + λ∑ Φ Di − ∑ Aij Di 

2
Ω
i ∈Ω j ∈ N (i )


First, the penalty function Φ(x) was adopted to act on the regular term, as
expressed below:
x
Φ (x) = − x
e +x
Φ(x) denotes the M-estimator. As reported by Smoli [28], robust M-estimators could
eliminate marginal uncorrelated patch. Figure 3 plots the L2 penalty function curve and
the penalty function curve applied in this study. The results indicated that the penalty
function proposed in this study was capable of efficiently saturating in the iteration, making
Φ(x) an approximation of L0 to keep the sharp

edges of the image. Compared with the
penalty function G(x) = 2 ∗ 1 − exp − x2 /2 applied in reference [29], the curve of Φ(x)
achieved a faster saturation rate, so it would be more sensitive to outliers (e.g., noises
and void).

Figure 3. The curves of different penalty functions: The penalty function Φ(x) proposed in this study
achieved the fastest saturation rate.

Second, to increase the reconstruction effect of color images with unclear edges, the
autoregressive coefficient Aij is expressed as:

1 D0  C I 
Aij = A F Ai,j , Ai,j
s i,j
0
D was defined by the Gaussian
where s denotes the normalization factor, the depth term Ai,j
filter, and the color term and the infrared term were, respectively, set as:
   2 
C h − ρh
∑
 h∈c i
B ◦ ρ i j 2 
C
Ai,j = exp
σ12


   2 
I h h
 ∑h∈c Bi
◦ ρi − ρ j

I 2
Ai,j = exp
σ12

Electronics 2022, 11, 993 6 of 22

where σ1 denotes the decay rate of the exponential function; ρih represents the mask operator
with the size ω ∗ ω centered on the pixel point within the color channel h. According to
the extracted operators, the bilateral filter BiC was adopted to adjust the distance and
intensity difference.
  
∑h∈c Cih − Cjh
!
2
k i − j k
BiC (i, j) = exp 2
exp− 
σ22 σ22

The filters BiC employed a bilateral kernel to weigh the distance of the local patch.
Compared with the Gaussian kernel adopted by the standard filter, the color term and the
infrared term showed a strong response to the pixels x with similar results. Accordingly,
the regression prediction of image edge pixels was significant.
Lastly, the check function F ( x ) was introduced in the regular term, as expressed below:
(
C , LBP (C ) ≥ LBP ( I )
Ai,j
F(x) = I , LBP (C ) ≺ LBP ( I )
Ai,j

To be specific, LBP( x ) represents an operator proposed by Ojala et al. [30] to describe

the local texture features of images; it is defined in 3 × 3 windows, while the value of
8 adjacent pixels was compared with the threshold value of the center pixel of the window.
If the surrounding pixel value was greater than the central pixel value, the position of the
pixel was marked as 1, but otherwise was 0. In this way, the LBP operator of the central
pixel of the window could be obtained, and this value could be used to reflect the texture
information of the region. The process was shown in Figure 4.

Figure 4. LBP operator definition.

In this study, the LBP equation is written as follows: LBP(C ) represents the ρ sampling
point in the circular area; R is the radius of the circular area; G (m) expresses the gray value
of the central pixel, and G ( p) denotes the gray value of the ρ circular boundary pixel.
ρ
LBPP,R ( xc , yc ) = ∑ s(G(ρ) − G(m)) ∗ 2ρ
n =1

Thus, in the regular term, the LBP function was adopted to judge the edge features
of the color image and the infrared image, a high-quality mask operator was selected
to act on the super-resolution model of the depth image. The effect of illumination and
distance on the depth image was solved effectively. The texture copy artifacts and the
depth discontinuities decreased significantly. Figure 5 gives the super-resolution result of
depth image.
Electronics 2022, 11, 993 7 of 22

Figure 5. Depth image restoration: (A) is the original depth image; (B) represents the reconstructed
depth image. Texture copy artifacts and depth discontinuities are significantly reduced.

In the real-world experiment, since there were no standard depth maps as reference,
Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) could not be used to
measure the quality of image reconstruction. In order to further verify the effectiveness
of this method, Mean Gradient (MG) and Laplace Operator (LO) [31] were selected as
evaluation indexes in this study. The higher MG and LO values were, the better image
quality was. The formula of MG was as follows:
1
MG = (M −1)( N −1)
∗
M −1 N −1
q
( F (i,j)− F (i +1,j))2 +( F (i,j)− F (i,j+1))2
∑ ∑ 2
i =1 J =1

In the formula, the gray value of row i and column j in the image was expressed as
F (i, j); M and N were the total number of rows and columns of the image. The encapsulation
method of Laplace Operator was provided in OpenCV, which could be called directly. The
formula of LO was as follows:
LO = ∑ ∑| G ( x, y)|
x y

The results were shown in Section 3.

2.3.2. Body-Part Segmentation

To improve the accuracy of the semantic segmentation in complex backgrounds and
reduce the segmentation error in up sampling, this study optimized the U-Net neural
network model based on PyTorch framework (Version 1.5.0) to achieve a remote, noncontact
and automatic detection of vital positions of cattle and goats, respectively. Figure 6 shows
the model.
Electronics 2022, 11, 993 8 of 22

Figure 6. Semantic segmentation models.

First, the U-Net neural network employed ReLU as the activation function. Since
the negative semi-axis value of ReLU was always zero, when the model inputted the
minimum value, the neuron would die because it was not activated. To solve the gradient
disappearance attributed to the ReLU activation function, the ReLU activation function after
each convolutional layer and Batch Normalization layer in the U-Net neural network was
replaced with the Mish activation function. The function of Mish and ReLU are illustrated
in Figure 7.

Figure 7. Curves for Mish and ReLU.

Since the positive semi-axis derivative of the Mish was greater than 1, it avoided the
saturation because of the function capping, while preventing the gradient from disappear-
ing. The negative semi-axis of the Mish was not always zero, whereas it remained a small
value, which solved the problem of neuron death; moreover, Mish improved the sparse
expression of the network and addressed the gradient disappearance attributed to the lack
of training data and the increase in network layers. According to existing studies, a smooth
activation function can more effectively allow the information to penetrate the neural
network [32]. Better generalization results can be achieved since the Mish is smoother than
that of ReLU.
 Electronics 2022, 11, 993 9 of 22
Electronics 2022, 11, x FOR PEER REVIEW 9 of 22

Second, this method introduced the attention mechanism to increase the accuracy
of semantic
suggested thatsegmentation. The attention
avoiding dimension mechanism
reduction helpsrefers to effective
learn a data processing
channelmethod
attention, and
in machine learning, originating from natural language processing. Relevant researchers
appropriate cross-channel interaction is capable of reducing model complexity while
suggested that avoiding dimension reduction helps learn effective channel attention, and
keeping performance
appropriate cross-channel[33].
interaction is capable of reducing model complexity while keep-
Accordingly,
ing performance the model could add the attention mechanism module after the respec-
[33].
tive down-sampled convolution
Accordingly, the model could addmodule. First,mechanism
the attention the binary adaptive
module mean
after the method was
respective
adopted to calculate
down-sampled and pool
convolution theFirst,
module. input thedata to adaptive
binary a size ofmean
1 × 1.method
After was
the global
adoptedaverage
to calculate
pooling, theand
1D pool the inputwith
convolution data weight of 1 × 1.characteristics
to a sizesharing After the globalwasaverage pooling,
exploited to capture
the 1D convolution with weight sharing characteristics was exploited to capture
the cross-channel interaction information. Next, the nonlinear relation was introduced the cross-
channel the
through interaction
sigmoidinformation. Next, theLastly,
activation function. nonlinear
the relation
result waswasmultiplied
introducedby through
elements and
the sigmoid activation function. Lastly, the result was multiplied by elements and then
then expanded into the same dimension as the input data, so the dimension of the data
expanded into the same dimension as the input data, so the dimension of the data processed
processed by themechanism
by the attention attention module
mechanism module
remained remained
unchanged. Byunchanged.
introducing theBy attention
introducing the
attention
module, themodule,
accuracythe accuracy
of semantic of semantic
segmentation segmentation
tasks tasks
was significantly was significantly
improved. Figure 8 im-
proved. Figure 8 presents the attention
presents the attention mechanism module. mechanism module.
1×1×C 1×1×C

Figure
Figure 8. ECA module.
ECA module.

Lastly, to achieve the desired segmentation effect more accurately and efficiently,
the method added the DropBlock2D module during each up sampling. The DropBlock
module [34] refers to a regularization module for the convolutional neural network. The
effect is illustrated in Figure 9. Such a module removed an area around a feature and forced
the network to learn other features of the object after multiple iterations of training.
Lastly, to achieve the desired segmentation effect more accurately and efficiently, the
method added the DropBlock2D module during each up sampling. The DropBlock mod-
ule [34] refers to a regularization module for the convolutional neural network. The effect
is illustrated in Figure 9. Such a module removed an area around a feature and forced the
network to learn other features of the object after multiple iterations of training.

Figure 9. DropBlock2D module.

After processing by the neural network model proposed in this study, the background
and body feature regions of cattle and goats could be effectively segmented. The segmenta-
tion effect is given in Figure 10.
Figure 9. DropBlock2D module.

After processing by the neural network model proposed in this study, the back-
ground and body feature regions of cattle and goats could be effectively segmented. The
segmentation effect is given in Figure 10.
Electronics 2022, 11, 993 10 of 22

Figure 10. Segmentation effect of improved neural network.

To further verify the effectiveness of the method, LabelImg was used to manually
annotate the images to obtain the real segmentation. Fifty images of cattle and goats were
randomly adopted from the test set, respectively, and the accuracy, the sensitivity, the
specificity and the dice similarity coefficient (dice) [35,36] were employed to assess to the
effect of the body-part segmentation method.

TP + TN
Accuracy =
TP + FP + TN + FN
TP
Sensitivity =
TP + FN
TN
Speci f icity =
TN + FP
2 ∗ TP
dice =
( TP + FN ) + ( TP + FP)
where TP represents the number of body-part pixels correctly identified as body-part
segmentation method compared to the real segmentation by LabelImg. FP represents
the number of body-part pixels correctly identified as Body-part segmentation method
compared to the pixels as non-body parts identified by LabelImg. TN represents the
number of body-part pixels incorrectly identified as the body-part segmentation method
compared to the pixels as non-body parts identified by LabelImg. FN denotes the number
of body-part pixels incorrectly identified as body-part segmentation method compared to
the real segmentation by LabelImg.
Since Dice is sensitive to internal filling, Hausdorff distance (HD) [37,38] is introduced
to evaluate the segmented boundary.
 

 


 


HD(A, M) = max max[d( ai , M )], max d m j , A
 i
 j 


 


whereA = { a1 , a2 , a3 . . . ai } represents the result point set segmented by automatic, and

M = m1 , m2 , m3 . . . m j represents the result point set segmented by manual. a and m
were points on the boundary respectively. d(x, y) is the Euclidean distance from x to y. The
results were shown in Section 3.

2.3.3. Feature Point Location

Feature point location refers to a vital research field in the computer vision. Over the
past few years, studies such as the human body joint point location [39–41] and the face key
point location [23,42,43] have shown that on the premise of sufficient data, deep learning
could extract better features compared with the manual construction of features. This
research solved the problem that traditional methods could not accurately locate livestock
feature points by created a multistage densely stacked hourglass network.
Electronics 2022, 11, 993 11 of 22

Dense Block
The stacked hourglass network comprises residual blocks. In this study, dense
blocks [44] were selected to replace residual blocks. Compared with residual blocks, dense
block networks exhibited a simple structure and fewer parameters than ResNet. Second,
dense blocks could send information between any two layers, thereby enhancing the reuse
of features and alleviating the problems of gradient vanishing and model degradation.
In a dense block, any two layers were connected to each other. The input of each
layer in the network was the output of all previous layers, and the feature map learned
by this layer would act as the input of the later network layer. The dense block structure
is expressed as xn = hn ([ x0 , x1 , x2 . . . xn−1 ]), where xn denotes the input of the layer n, [ ]
represents splicing, combining all output feature graphs from x0 to xn−1 layers according
to channels, and hn expresses a composite function (i.e., Batch Norm, activation function
and a 3 ∗ 3 convolution).

Objective Function Optimization

The stacked hourglass network used the mean square error loss (L2 loss function) as
the loss function, and the calculation equation is written as: S = ∑in=1 (Yi − f ( xi ))2 . The
L2 loss function caused the model to produce a greater loss at the noise point. When the
model optimized the loss of noise points, the performance of the model would be reduced;
thus, this study adopted the Huber loss function to replace the original loss function to
reduce the sensitivity of the network to noise points and endow the network with a better
generalization capability. The Huber loss function equation is presented below:
(
1 2
Lδ (y, f ( x )) = 2 ( y − f ( x ))
1 2
δ|y − f ( x )| − 2 δ

The Huber loss is essentially an absolute error, only becoming a squared error when
the error was small. The hyper parameter δ was used to control when it became a quadratic
error. When the value of the Huber loss function was in the interval [0 − δ, 0 + δ], the
Huber loss function was equal to mean square error. When the value of Huber loss function
was in the interval [−∞, δ] or the interval [δ, + ∞], the Huber loss function was equated
with Mean Absolute Error. Accordingly, the Huber loss function combined the advantages
of MSE and MAE, and it would be more robust to abnormal points.

Multistage Densely Stacked Hourglass Network

Accurate positioning of the pixel position of the feature points of the livestock is of
great significance to improve the accuracy of the livestock size measurement. In recent
years, deep learning technology has been widely used, but there was little research on
feature point location of livestock. Referred to the idea of human feature points location
based on deep learning [45,46], this study used multistage densely stacked hourglass
network to locate the feature points of livestock.
The stacked hourglass network mainly consisted of hourglass sub-networks, and
Figure 11 illustrates the hourglass network. The upper and lower halves of the graph
contain several light green dense blocks to gradually extract deeper features. Dense blocks
were mainly composed of residual modules. The step and padding of all convolution
layers for all residual modules were set to 1. All operations did not change the data size;
they only changed the data depth (channel). The value C_IN in the first row of the dense
block represents the number of channels of the input image, and that in the second row
represents the number of output channels. The red module represents the down sampling
layer, using the max pooling method. The gray module represents the upper sampling
layer, using the nearest neighbor interpolation method. The dashed box could increase the
order of the network by splicing other hourglass sub-networks.
Electronics 2022, 11, 993 12 of 22

Figure 11. The structure of hourglass sub-networks.

The multistage densely stacked hourglass network refers to a serial stack of hourglass
sub-networks; however, with the increase in the network depth, the parameters increased,
making it difficult to perform effective training updates. To solve such a problem, this
study adopted the relay supervision strategy in the model for the supervised training. The
experiments in this study were all performed on a three-stage stacked hourglass network.
Figure 12 gives the four-level densely stacked hourglass network.

Figure 12. The structure of the three-level densely stacked hourglass network.

After the original image was preprocessed, the scale information of the image was
preserved for the next layer of feature fusion. In addition, the preprocessed image was
input into the hourglass sub-networks for a down sampling operation. After three residual
modules and two convolution modules, the features were extracted step by step, and the
feature map of the layer was obtained through the up-sampling operation. At the same
time, after up sampling operation, the scale information of the upper layer was fused and
used as the input of the next hourglass sub-networks to extract the features.
This study adopted the PyTorch framework for training, the input image was a color
image with a resolution of 768 × 768, and the number of iterations was set to 16,000.
The training set, validation set, and test set were divided according to the proportion of
5:3:2. For every 10,000 steps, the learning rate declined to 10%. Moreover, to increase the
convergence speed, the normalization operation was added to the model.

2.4. Evaluation of Body Size Measurement

Withers height, hip height, body length and chest depth of 13 cattle and 10 goats
were measured manually and automatically. The absolute error, relative error, linear
regression and repeatability have been used to verify the validity of the body size measure-
ment method.
First, the absolute error was used to reflect the deviation between automatic mea-
surement and manual measurement. Since the absolute error and the measured value
have same unit, the relative error was introduced to compare the reliability of the
automatic measurement.
Second, the ideal relationship between manual measurement and automatic mea-
surement should be Y = AX while A = 1. This research used linear regression to fit the
Electronics 2022, 11, 993 13 of 22

relationship between manual measurement and automatic measurement. The effectiveness

of the measurement method was verified by coefficient of determination and RMSE. A high
coefficient of determination R2 determination represented a good fit. RMSE was used to
reveal the typical distance between predicted value made by the regression model and the
actual value.
Finally, for the repeatability evaluation of the measurement method, the researchers
often used the intraclass correlation efficient (ICC) to assess the similarity of a quantitative
attribute among individuals and to assess different measurement methods or assess the re-
peatability of the same measurement result by the observer [47]. Measurement experiments
generally employed ICC to assess the repeatability of different measurement methods to
the same target [12]. The precondition of this study was that the manual measurement was
considered to be the most reliable and repeatable; thus, we just assessed the repeatability of
automatic measurement to the target by ICC and Coefficient of Variation (CV).

3. Results and Discussion

3.1. Depth Image Restoration
Ten depth images of cattle and goats and their corresponding color images and infrared
images were randomly selected from the data set. Joint Bilateral Upsampling (JBU) [48],
Weighted Least Squares (WLS) [49], Autoregressive Model (AR) [50] and the algorithm
proposed in this research were used to reconstruct the depth images, respectively. The
results are shown in Table 1.

Table 1. Comparison of MG and LO values of reconstruction results.

Goat Cattle
MG LO MG LO
JBU 0.0027 22.646 0.0028 21.404
WLS 0.0029 21.967 0.0029 22.142
AR 0.0032 74.482 0.0032 86.814
Ours 0.0048 142.170 0.0050 154.748

The results showed that the depth image reconstruction effect of AR algorithm was
better than JBU algorithm and WLS algorithm, but there were still deep texture copy
artifacts in the reconstructed depth image. In this research, making full use of the texture
features of infrared image could greatly improve the reconstruction quality of depth images.
The effect of illumination and distance on the depth image was solved effectively. The
texture copy artifacts and the depth discontinuities decreased significantly. Compared with
other algorithms, our method improved both visual effect and image quality.

3.2. Body-Part Segmentation

The cattle and goat body segmentation contains front trunk segmentation, central body
segmentation and hip–torso segmentation. The average of accuracy reached 93.68% and
93.50%, the sensitivity was 71.65% and 68.90%, and the specificity was 97.78% and 97.97%,
respectively (Tables 2 and 3). For background segmentation of cattle and goat, the accuracy
reached 97.09% and 96.73%, the sensitivity was 89.29% and 90.00%, the specificity was
99.57% and 99.40%, and the dice similarity coefficient was 93.67% and 92.99%, respectively
(Tables 2 and 3). The Hausdorff distances of background segmentation between cattle
and goat were 5.93 mm and 6.16 mm, respectively (Table 4). Compared to background
segmentation, body segmentation required further segmentation in a smaller segmentation
area and a more similar color background, which might lead to a relatively worse effect of
body segmentation than background segmentation. For body segmentation, the hip torso
was a significantly poor segmentation effect compared to front trunk segmentation and
central body segmentation. As impacted by the low sensitivity of the segmentation method
to the tails of cattle and goats, the tails and trunks of livestock could not be separated from
the background.
Electronics 2022, 11, 993 14 of 22

Table 2. Body-part detection performance parameters for side view of cattle (%) *.

Index Background Front Trunk Central Body Hip Torso p-Value

97.09 ± 1.56 a b ab 89.55 ± 1.29 c
Accuracy 94.30 ± 1.53 95.33 ± 1.70 <0.001
Sensitivity 89.29 ± 1.11 a 54.55 ± 1.63 d 80.00 ± 1.26 b 71.43 ± 1.02 c <0.001
Specificity 99.57 ± 0.60 a 99.75 ± 0.60 a 97.83 ± 0.57 b 94.34 ± 0.64 c <0.001
Dice 93.67 ± 0.71 a 69.76 ± 0.66 b 82.75 ± 0.54 c 74.07 ± 0.79 d <0.001
* All data were presented as average ± SD. One-way analysis of variance (ANOVA) followed by Turkey’s multiple-
range test was used to determine significant differences among different segmentations. Different superscript
letters indicate significant differences (p < 0.05).

Table 3. Body-part detection performance parameters for side view of goats (%) *.

Index Background Front Trunk Central Body Hip Torso p-Value

96.73 ± 0.95 a b b 87.59 ± 1.16 c
Accuracy 95.27 ± 0.85 95.52 ± 1.09 <0.001
Sensitivity 90.00 ± 1.39 a 60.98 ± 1.29 c 76.92 ± 1.15 b 60.81 ± 1.24 d <0.001
Specificity 99.40 ± 0.62 a 99.26 ± 0.61 a 98.23 ± 0.63 b 95.13 ± 0.69 c <0.001
Dice 92.99 ± 0.89 a 72.88 ± 0.83 b 81.38 ± 0.49 c 68.28 ± 0.91 d <0.001
* All data were presented as average ± SD. One-way analysis of variance (ANOVA) followed by Turkey’s multiple-
range test was used to determine significant differences among different segmentations. Different superscript
letters indicate significant differences (p < 0.05).

Table 4. Hausdorff distance of body segmentation between cattle and goat (mm).

Index Cattle Goat

Background 5.93 ± 1.38 6.16 ± 1.41
front trunk 6.74 ± 1.56 6.94 ± 1.49
central body 6.23 ± 1.12 6.78 ± 1.23
Hip torso 7.37 ± 1.74 8.40 ± 1.77

3.3. Accuracy of Feature Point Location

This study selected 2, 4, 6, 8 and 10 dense hourglass blocks for testing. Figure 13 shows
the experimental results. To be specific, a–e represents the quantitative feature points of
cattle and goats marked in Figure 2.

Figure 13. Feature point location results of different depth networks.

Evaluation used the percentage of correct key points (σ = 0.5) [39,51] metric which
reports the percentage of detections that fall within a normalized distance of the ground
truth. In this study, because some pictures did not collect the head information of livestock,
Electronics 2022, 11, 993 15 of 22

the body length of livestock was selected to calculate the normalized distance between the
predicted feature point and the real feature point.
This study used the self-built data set to train the multistage densely stacked hourglass
network. The evaluation reported that the model exhibited high universality and could
accurately locate the feature points of cattle and goats. Figure 14 presents the location effect.

Figure 14. (A): The blue points are the manually marked feature points and the orange points are the
feature points identified by the algorithm. From left to right and from top to bottom of the right half
of Figure 14A is an enlarged image of the corresponding positions of points a, d, e, b and c in (B):
Body size measurement results.

3.4. Accuracy of Measurement

According to Figures 15 and 16, most automatic measurement results were greater
than the manual measurement results. Given the Biltmore stick clings to livestock torso and
less affected by the livestock fur during manual measurements, the livestock fur greatly
affected the background segmentation and the feature point location. The method located
feature points were close to fur of the livestock, but not close to body. The result of the
automatic measurement was slightly greater than that of the manual measurement.
The box plots of the absolute error and relative error between manual measurement
and automatic measurement were shown in Figures 17 and 18. The average absolute error
of withers height, hip height, and chest depth were mostly between 1 and 2 cm, and most
accuracies of withers height were greater than hip height, body length and chest depth.
The body length had the largest average absolute error, which may because the shoulder
point (point b in Figure 2) was located at the front edge of the shoulder, greatly affected by
the animal’s posture. Moreover, the color and depth of the feature points were similar to
Electronics 2022, 11, 993 16 of 22

those of the surrounding pixel points, resulted the great error in body length measurement.
The mean relative errors of withers height, hip height, body length and chest depth were
0.68%, 0.68%, 1.39% and 1.60% for cattle, and 0.65%, 1.06%, 1.77% and 1.99% for goat,
respectively. The average relative error of the measurement results was between −2% to
2%; therefore, the results of the automatic measurement could be trustworthy in terms of
the measurement accuracy.

Figure 15. Comparison of manual measurement and the proposed method for the four different cattle
body parameters: (A) Withers height, (B) hip height, (C) body length and (D) chest depth.

Figure 16. Comparison of manual measurement and the proposed method for the four different goat
body parameters: (A) Withers height, (B) hip height, (C) body length and (D) chest depth.
Electronics 2022, 11, 993 17 of 22

Figure 17. Box plots of the absolute error: (A) Box plots of the absolute error of the measurement of
the cattle body size; (B) Box plots of the absolute error of the measurement of the goat body size.

Figure 18. Box plots of relative error: (A) Box plots of the relative error the measurement of the cattle
body size; (B) Box plots of the relative error the measurement of the goat body size.

3.5. The Linear Regerssion Analysis of Measurement Results

As shown in Figure 19, the coefficient of determination of withers height, hip height
body length and chest depth of cattle were 0.9974, 0.9985, 0.9954 and 0.9579, and RMSE were
2.492, 2.939, 3.328 and 1.927, which demonstrated that the two variables had a high degree
of explanation. The linear regression coefficient of the cattle body size measured was close to
1, which showed the consistency of automatic measurement with the manual measurement.
As shown in Figure 20, the coefficient of determination of withers height, hip height
body length and chest depth of goats were 0.9974, 0.9985, 0.9954 and 0.9579, and RMSE were
1.289, 1.419, 1.932 and 0.824. The coefficient of determination of automatic measurement
was slightly lower than cows. Since goats are smaller than cattle, it is difficult to locate
the feature points. In fact, whether it is automatic measurement or manual measurement,
when the size of livestock decreases or the bone is unclear, the accuracy of feature point
location will decrease [52]; however, the linear regression coefficient of the goats’ body size
measured was close to 1, which showed the consistency of two measurement methods.
Electronics 2022, 11, 993 18 of 22

Figure 19. Linear Regression Analysis of cattle body measurement.

Figure 20. Linear Regression Analysis of goat body measurement.

3.6. Repeatability of Measurement Results

Table 5 lists the repeatability and reproducibility analysis results of this study after ten
measurements. The value of ICC was between zeros to one. If the value of ICC was less
than 0.4, the repeatability of the experiment would be poor; if the value of ICC was greater
than 0.75, the experiment would exhibit better repeatability. In this study, the ICC of all
Electronics 2022, 11, 993 19 of 22

body measurements was greater than 0.9, thereby showing good measurement repeatability.
The ICC of withers height and hip height of cattle and withers height of goats all exceeded
0.980, which demonstrated higher reliability than other. Using the automatic measurement
to measure body size of cattle, the value of CV ranged from 0.049% to 0.090% (variance
ranges from 6.91 cm to 10.83 cm). The range of CV for goat body measurements ranged
from 0.043% to 0.141% (variance ranges from 2.79 cm to 4.43 cm). According to Fischer,
measurement methods with the repeatability and reproducibility had a CV less than 4%, as
in the present study were considered promising [53].

Table 5. The analysis results of repeatability and reproducibility *.

ICC CV
Cattle Goat
Cattle Goat
Manual Automatic Manual Automatic
withers height 0.990 0.985 0.064% 0.065% 0.052% 0.054%
hip height 0.985 0.943 0.052% 0.049% 0.041% 0.043%
body length 0.973 0.890 0.068% 0.069% 0.056% 0.064%
chest depth 0.972 0.942 0.091% 0.090% 0.123% 0.141%
* ICC describes the reliability and repeatability of measurements for clusters. The value of ICC close to 1 indicates
high reliability and repeatability. CV is a normalized measure of the dis-persion of probability distribution, which
is defined as the ratio of standard deviation to average.

Table 5 lists the body size measurement result of cattle, better than that of goats.
Besides the effect of fur, another reason was that during the depth data collection, the
behaviors of goats were more active than cattle; therefore, the distance between livestock
and depth camera is not consistent. In many goat images, the area occupied by the torso is
smaller than that of cattle; thus, there were more holes and texture copying artifacts in the
depth images. The mentioned holes and texture replication artifacts would exist around
the feature points, and would also affect the background segmentation and feature point
extraction even after the depth image super-resolution reconstruction.

4. Conclusions and Future Studies

This study proposed a livestock body size measurement method and could measure
various livestock body sizes under different backgrounds, respectively. The innovation of
this method mainly includes the following:
(1) Based on autoregressive models, a new penalty function was used to complete depth
image super-resolution reconstruction through the color image and infrared image as
guide image. The effect of illumination, distance and other factors on the quality of
depth images was reduced.
(2) Combining the characteristics exhibited by the attention module and the DropBlock2D
module, this study used a self-built database to realize remote, contactless, automatic
background segmentation and trunk segmentation of cattle and goats, respectively,
by optimizing the activation function of U-Net neural network.
(3) It was the first time to use deep learning technology to locate livestock feature points.
The three-stage stacked hourglass network was constructed by optimizing the objec-
tive function;l moreover, the relay supervision strategy was applied for the supervised
training to achieve the feature point location of cattle and goats.
The results showed that the average accuracies of cattle and goat body segmentation
were 93.68% and 93.50%, the sensitivities were 71.65% and 68.90%, and the specificities
were 97.78% and 97.97%, respectively. The percentage of correct key points for the detection
of the points of the withers, the shoulder points, the shallowest part of the chest, the highest
point of the hip bones and the ischia tuberosity took up 96.7%, 89.3%, 95.6%, 90.5% and
94.5%, respectively. Compared with the manual measurement, the accuracy, correlation
Electronics 2022, 11, 993 20 of 22

and repeatability of the measurement of with height, hip height body length and chest
depth were better.
This study used deep learning to estimate the four body measurements of cattle and
goat, and it achieved good results; it proved that the deep learning method could be
applied to livestock trunk segmentation, body feature point location, and livestock body
size measurement. The future research is mainly divided into two parts. First, due to the
limitation of the number of data set, the current research object was limited to cattle and
goat. In future research, more types of livestock will be collected and annotated to achieve
the body measurement of different kinds of livestock (even horses and donkeys) with the
same feature point. Second, only the side views of cattle and goat have been studied. When
the side view of livestock cannot be collected completely, the measurement accuracy will
be reduced; therefore, more angles of livestock images (top view or images from different
angles) will be added in future research to further study the normalization of livestock
posture and measure more livestock body sizes. Finally, as the preliminary research, only
image processing was supported at present. In future research, a video stream will be used
to replace images to complete the measurement of livestock.

Author Contributions: Methodology, K.L.; validation, K.L.; resources, G.T.; writing—original draft
preparation, K.L. All authors have read and agreed to the published version of the manuscript.
Funding: This work was supported by the National Natural Science Foundation of China (Grant No.
U20A20180), the Graduate Innovation Foundation of Hebei (Grant No. CXZZBS2021045).
Data Availability Statement: This experiment does not involve a livestock slaughter experiment,
and the data collection process of related livestock conformed to the regulations of the animal Ethics
Committee of Hebei Agricultural University. And this study not involving humans.
Conflicts of Interest: The authors declare that they have no known competing financial interests or
personal relationships that could have appeared to influence the work reported in this paper.

References
1. Thorup, V.M.; Edwards, D.; Friggens, N.C. On-farm estimation of energy balance in dairy cows using only frequent body weight
measurements and body condition score. J. Dairy Sci. 2012, 95, 1784–1793. [CrossRef] [PubMed]
2. Pezzuolo, A.; Guarino, M.; Sartori, L.; González, L.A.; Marinello, F. On-barn pig weight estimation based on body measurements
by a Kinect v1 depth camera. Comput. Electron. Agric. 2018, 148, 29–36. [CrossRef]
3. Kuzuhara, Y.; Kawamura, K.; Yoshitoshi, R.; Tamaki, T.; Sugai, S.; Ikegami, M.; Kurokawa, Y.; Obitsu, T.; Okita, M.; Sugino, T. A
preliminarily study for predicting body weight and milk properties in lactating Holstein cows using a three-dimensional camera
system. Comput. Electron. Agric. 2015, 111, 186–193. [CrossRef]
4. Menesatti, P.; Costa, C.; Antonucci, F.; Steri, R.; Pallottino, F.; Catillo, G. A low-cost stereovision system to estimate size and
weight of live sheep. Comput. Electron. Agric. 2014, 103, 33–38. [CrossRef]
5. Leonard, S.M.; Xin, H.; Brown-Brandl, T.M.; Ramirez, B.C. Development and application of an image acquisition system for
characterizing sow behaviors in farrowing stalls. Comput. Electron. Agric. 2019, 163, 104866. [CrossRef]
6. Brandl, N.; Rgensen, E. Determination of live weight of pigs from dimensions measured using image analysis. Comput. Electron.
Agric. 1996, 15, 57–72. [CrossRef]
7. Marchant, J.A.; Schofield, C.P.; White, R.P. Pig growth and conformation monitoring using image analysis. Anim. Sci. 1999, 68,
141–150. [CrossRef]
8. Tasdemir, S.; Urkmez, A.; Inal, S. Determination of body measurements on the Holstein cows using digital image analysis and
estimation of live weight with regression analysis. Comput. Electron. Agric. 2011, 76, 189–197. [CrossRef]
9. Ozkaya, S. Accuracy of body measurements using digital image analysis in female Holstein calves. Anim. Prod. Sci. 2012, 52,
917–920. [CrossRef]
10. Federico, P.; Roberto, S.; Paolo, M.; Francesca, A.; Corrado, C.; Simone, F.; Catillo, G. Comparison between manual and
stereovision body traits measurements of Lipizzan horses. Comput. Electron. Agric. 2015, 118, 408–413.
11. Shi, C.; Teng, G.H.; Li, Z. An approach of pig weight estimation using binocular stereo system based on LabVIEW. Comput.
Electron. Agric. 2016, 129, 37–43. [CrossRef]
12. Salau, J.; Haas, J.H.; Junge, W.; Thaller, G. A multi-Kinect cow scanning system: Calculating linear traits from manually marked
recordings of Holstein-Friesian dairy cows. Biosyst. Eng. 2017, 157, 92–98. [CrossRef]
13. Wang, K.; Guo, H.; Ma, Q.; Su, W.; Zhu, D.H. A portable and automatic Xtion-based measurement system for pig body size.
Comput. Electron. Agric. 2018, 148, 291–298. [CrossRef]
Electronics 2022, 11, 993 21 of 22

14. Shi, S.; Yin, L.; Liang, S.H.; Zhong, H.J.; Tian, X.H.; Liu, C.X.; Sun, A.; Liu, H.X. Research on 3D surface reconstruction and body
size measurement of pigs based on multi-view RGB-D cameras. Comput. Electron. Agric. 2020, 175, 105543. [CrossRef]
15. Ruchay, A.; Kober, V.; Dorofeev, K.; Kolpakov, V.; Miroshnikov, S. Accurate body measurement of live cattle using three depth
cameras and non-rigid 3-D shape recovery. Comput. Electron. Agric. 2020, 179, 105821. [CrossRef]
16. Lina, Z.A.; Pei, W.B.; Tana, W.C.; Xinhua, J.D.; Chuanzhong, X.E.; Yanhua, M.F. Algorithm of body dimension measurement and
its applications based on image analysis. Comput. Electron. Agric. 2018, 153, 33–45. [CrossRef]
17. Guo, H.; Ma, X.; Ma, Q.; Wang, K.; Su, W.; Zhu, D.H. LSSA_CAU: An interactive 3d point clouds analysis software for body
measurement of livestock with similar forms of cows or pigs. Comput. Electron. Agric. 2017, 138, 60–68. [CrossRef]
18. Guo, H.; Li, Z.B.; Ma, Q.; Zhu, D.H.; Su, W.; Wang, K.; Marinello, F. A bilateral symmetry based pose normalization framework
applied to livestock body measurement in point clouds. Comput. Electron. Agric. 2019, 160, 59–70. [CrossRef]
19. Zhao, K.X.; Li, G.Q.; He, D.J. Fine Segment Method of Cows’ Body Parts in Depth Images Based on Machine Learning. Nongye
Jixie Xuebao 2017, 48, 173–179.
20. Jiang, B.; Wu, Q.; Yin, X.Q.; Wu, D.H.; Song, H.B.; He, D.J. FLYOLOv3 deep learning for key parts of dairy cow body detection.
Comput. Electron. Agric. 2019, 166, 104982. [CrossRef]
21. Li, B.; Liu, L.S.; Shen, M.X.; Sun, Y.W.; Lu, M.Z. Group-housed pig detection in video surveillance of overhead views using
multi-feature template matching. Biosyst. Eng. 2019, 181, 28–39. [CrossRef]
22. Song, X.; Bokkers, E.; Mourik, S.V.; Koerkamp, P.; Tol, P. Automated body condition scoring of dairy cows using 3-dimensional
feature extraction from multiple body regions. J. Dairy Sci. 2019, 102, 4294–4308. [CrossRef] [PubMed]
23. Zhang, J.; Shan, S.G.; Kan, M.; Chen, X.L. Coarse-to-Fine Auto-Encoder Networks (CFAN) for Real-Time Face Alignment. In
Proceedings of the European Conference on Computer Vision, Zurich, Switzerland, 6–12 September 2014; pp. 1–16.
24. Cozler, Y.L.; Allain, C.; Xavier, C.; Depuille, L.; Faverdin, P. Volume and surface area of Holstein dairy cows calculated from
complete 3D shapes acquired using a high-precision scanning system: Interest for body weight estimation. Comput. Electron.
Agric. 2019, 165, 104977. [CrossRef]
25. Wang, K.; Zhu, D.; Gu, O.H.; Ma, Q.; Su, W.; Su, Y. Automated calculation of heart girth measurement in pigs using body surface
point clouds. Comput. Electron. Agric. 2019, 156, 565–573. [CrossRef]
26. Weisheng, D.; Lei, Z.; Rastislav, L.; Guangming, S. Sparse representation based image interpolation with nonlocal autoregressive
modeling. IEEE Trans. Image Process. 2013, 22, 1382–1394.
27. Hornácek, M.; Rhemann, C.; Gelautz, M.; Rother, C. Depth Super Resolution by Rigid Body Self-Similarity in 3D. In Proceedings
of the 2013 IEEE Conference on Computer Vision and Pattern Recognition, Portland, OR, USA, 23–28 June 2013; pp. 1123–1130.
[CrossRef]
28. Smoli, A.; Ohm, J.R. Robust Global Motion Estimation Using A Simplified M-Estimator Approach. In Proceedings of the 2000
International Conference on Image Processing, Vancouver, BC, Canada, 10–13 September 2000; pp. 868–871.
29. Zhu, R.; Yu, S.J.; Xu, X.Y.; Yu, L. Dynamic Guidance for Depth Map Restoration. In Proceedings of the 2019 IEEE 21st International
Workshop on Multimedia Signal Processing (MMSP), Kuala Lumpur, Malaysia, 27–29 September 2019; pp. 1–6. [CrossRef]
30. Ojala, T.; Pietikainen, M.; Maenpaa, T. Multiresolution Gray-Scale and Rotation Invariant Texture Classification with Local Binary
Patterns. IEEE Trans. Pattern Anal. Mach. Intell. 2002, 24, 971–987. [CrossRef]
31. Said, P.; Domenec, P.; Miguel, A.G. Analysis of focus measure operators for shape-from-focus. Pattern Recognit. 2013, 46,
1415–1432.
32. Misra, D. Mish: A Self Regularized Non-Monotonic Neural Activation Function. In Proceedings of the British Machine Vision
Conference, Cardiff, UK, 9–12 September 2019.
33. Wang, Q.; Wu, B.; Zhu, P.; Li, P.; Zuo, W.; Hu, Q. ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks.
In Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA, 14–19
June 2020; pp. 11531–11539.
34. Ghiasi, G.; Lin, T.Y.; Le, Q.V. Dropblock: A regularization method for convolutional networks. arXiv 2018, arXiv:1810.12890.
35. Shibata, E.; Takao, H.; Amemiya, S.; Ohtomo, K. 3D-Printed Visceral Aneurysm Models Based on CT Data for Simulations of
Endovascular Embolization: Evaluation of Size and Shape Accuracy. Am. J. Roentgenol. 2017, 209, 243–247. [CrossRef]
36. Carass, A.; Roy, S.; Gherman, A.; Reinhold, J.C.; Jesson, A.; Arbel, T.; Maier, O.; Handels, H.; Ghafoorian, M.; Platel, B.; et al.
Evaluating White Matter Lesion Segmentations with Refined Sørensen-Dice Analysis. Sci. Rep. 2020, 10, 8242. [CrossRef]
37. Abdel Aziz, T.; Allan, H. An Efficient Algorithm for Calculating the Exact Hausdorff Distance. IEEE Trans. Pattern Anal. Mach.
Intell. 2015, 37, 2153–2163.
38. Karimi, D.; Salcudean, S.E. Reducing the Hausdorff Distance in Medical Image Segmentation with Convolutional Neural
Networks. IEEE Trans. Med. Imaging 2020, 39, 499–513. [CrossRef] [PubMed]
39. Newell, A.; Yang, K.; Deng, J. Stacked Hourglass Networks for Human Pose Estimation. In Proceedings of the European
Conference on Computer Vision, Amsterdam, The Netherlands, 8–16 October 2016; pp. 483–499.
40. Chu, X.; Yang, W.; Ouyang, W.; Ma, C.; Yuille, A.L.; Wang, X.G. Multi-Context Attention for Human Pose Estimation. In
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA, 21–26 July 2017;
pp. 5669–5678.
41. Cao, Z.; Simon, T.; Wei, S.E.; Sheikh, Y. Realtime Multi-person 2D Pose Estimation Using Part Affinity Fields. In Proceedings of
the 2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA, 21–26 July 2017; pp. 1302–1310.
Electronics 2022, 11, 993 22 of 22

42. Sun, Y.; Wang, X.G.; Tang, X.O. Deep Convolutional Network Cascade for Facial Point Detection. In Proceedings of the 2013 IEEE
Conference on Computer Vision and Pattern Recognition, Portland, OR, USA, 23–28 June 2013; pp. 3476–3483.
43. Ranjan, R.; Patel, V.M.; Chellappa, R. HyperFace: A Deep Multi-task Learning Framework for Face Detection, Landmark
Localization, Pose Estimation, and Gender Recognition. IEEE Trans. Pattern Anal. Mach. Intell. 2019, 41, 121–135. [CrossRef]
[PubMed]
44. Huang, G.; Liu, Z.; Laurens, V.; Weinberger, K.Q. Densely Connected Convolutional Networks. In Proceedings of the 2017 IEEE
Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA, 21–26 July 2017; pp. 2261–2269.
45. Hua, G.; Li, L.; Liu, S. Multipath affinage stacked-hourglass networks for human pose estimation. Front. Comput. Sci. 2020, 14,
155–165. [CrossRef]
46. Bao, W.; Yang, Y.; Liang, D.; Zhu, M. Multi-Residual Module Stacked Hourglass Networks for Human Pose Estimation. J. Beijing
Inst. Technol. 2020, 29, 110–119.
47. Donner, A.; Koval, J.J. The estimation of intraclass correlation in the analysis of family data. Biometrics 1980, 36, 19–25. [CrossRef]
48. Johannes, K.; Michael, F.C.; Dani, L.; Matt, U. Joint bilateral upsampling. ACM Trans. Graph. 2007, 26, 96.1–96.4.
49. Camplani, M.; Mantecón, T.; Salgado, L. Depth-Color Fusion Strategy for 3-D Scene Modeling With Kinect. IEEE Trans. Cybern.
2013, 43, 1560–1571. [CrossRef]
50. Yang, J.; Ye, X.; Li, K.; Hou, C.; Wang, Y. Color-guided depth recovery from RGB-D data using an adaptive autoregressive model.
IEEE Trans. Image Process. 2014, 23, 3443–3458. [CrossRef]
51. Enwei, S.; Ramakrishna, V.; Kanade, T.; Sheikh, Y. Convolutional pose machines. In Proceedings of the IEEE Computer Society,
Las Vegas, NV, USA, 27–30 June 2016; pp. 4724–4732.
52. Cozler, Y.L.; Allain, C.; Caillot, A.; Delouard, J.M.; Delattre, L.; Luginbuhl, T.; Faverdin, P. High-precision scanning system
for complete 3D cow body shape imaging and analysis of morphological traits. Comput. Electron. Agric. 2019, 157, 447–453.
[CrossRef]
53. Fischer, A.; Luginbühl, T.; Delattre, L.; Delouard, J.M.; Faverdin, P. Rear shape in 3 dimensions summarized by principal
component analysis is a good predictor of body condition score in Holstein dairy cows. J. Dairy Sci. 2015, 98, 4465–4476.
[CrossRef] [PubMed]