Texture-based Fruit Detection via Images

using the Smooth Patterns on the Fruit

Zania S Pothen1 and Stephen Nuske2
Abstract This paper describes a keypoint detection algorithm to accurately detect round fruits in high resolution
imagery. The significant challenge associated with round fruits
such as grapes and apples is that the surface is smooth and lacks
definition and contrasting features, the contours of the fruit may
be partially occluded, and the color of the fruit often blends
with background foliage. We propose a fruit detection algorithm
that utilizes the gradual variation of intensity and gradient
orientation on the surface of the fruit. Candidate fruit locations,
or seed points are tested for both monotonically decreasing
intensity and gradient orientation profiles. Candidate fruit
locations that pass the initial filter are classified using modified
histogram of oriented gradients combined with a pairwise
intensity comparison texture descriptor and random forest
classifier. We analyse the performance of the fruit detection
algorithm on image datasets of grapes and apples using human
labeled images as ground truth. Our method to detect candidate
fruit locations is scale invariant, robust to partial occlusions
and more accurate than existing methods. We achieve overall
F1 accuracy score of 0.82 for grapes and 0.80 for apples.
We demonstrate our method is more accurate than existing
methods.

(a) Imaging Unit in Vineyard

I. I NTRODUCTION
Growers of fruit crops currently have access to very
limited information from the current state of the crop in
their field. Automated image analysis systems for fruit crops
provide growers with high resolution spatial data on their
crop yield. This data enables farmers to switch from old
inefficient farming one-size fits all paradigm to a more
cost-effective, site-specific field management system boosting the crop quality and conserving resources on the farm.
Such automated imaging systems can also be integrated into
fully robotic harvesting or thinning systems.
Currently there are no systems available for growers to
measure crop yield with high resolution during the growing
season. Crop yield is a desirable attribute to be monitored
and managed. The current process to estimate yield is by
monitoring the farm at harvest time, and recording data
during each growing season. However, yield can vary by
large amounts from year to year, and using harvest estimates
is an extremely coarse approximation of yield. In order to
get accurate dense measures of crop yield, the crop needs to
be continuously measured during the growing season. The
obvious solution would be to exhaustively monitor the fields.
*This work supported by the United States Department of Agriculture
(Award number: 2012-67021-19958) and the National Grape and Wine
Initiative ([email protected]).
1 Zania S Pothen, Carnegie Mellon University, 5000 Forbes Avenue,
Pittsburgh, PA 15213 [email protected]
2 Stephen Nuske, Carnegie Mellon University, 5000 Forbes Avenue,
Pittsburgh, PA 15213 [email protected]

(b) grape

(c) apple

(d) partially occluded apple

Fig. 1: An overview of the Yield Mapping System (a) Imaging Unit in
Vineyard consisting of stereo cameras and a pair of flashes for capturing
high resolution images. [(b),(c),(d)] Example images of smooth round fruit
and the distinct patterns formed on them - (b) a grape, (c) an apple and (d)
the visible portion of a partially occluded apple. (Left) RGB Images of the
fruit. (Middle) The Instensity pattern on the fruit surface viewed using the
jet heatmap of the grayscale image. The deep red tones represent intensity
values close to the maximum intensity found at the center of the fruit while
the successive color tones represent intensity bands of decreasing strength.
(Right) The Orientation pattern viewed using a jet heatmap of the gradient
orientation image of the fruit.

While the approach might work well for small sized fields,
it becomes economically intractable for larger fields owing
to the labour intensive nature of the work. Additionally, the
manual counting mechanism is performed just before harvest.
Over the past few years, our research group has focused
on developing a vision-based system for automatic fruitdetection and high resolution yield-estimation. Our current
system is deployed on a vehicle operating at high velocities
(>1.5 m/s), and captures images using a custom hardware
configuration and high-powered flash lighting. The images
are processed for crop-yield statistics.
The broad steps employed in predicting yield automatically and non-destructively are:
Collect images of fruit-wall in each row of the field
using custom hardware (Figure 1a)
Detect and count fruit in individual images
Associate data between fruit locations in the image and
physical locations in the real-world
Generate high resolution yield estimates using the registered fruit information and the vehicle state information
The details of the system and approach is described in Nuske
et al. [12][14]
We are motivated to increase accuracy of our system and
this paper focuses on novel image processing approaches
to specifically leverage the unique smooth patterns on the
surface of the fruit (Figure 1(b-d)). The smooth texture on
the fruits surface results in a distinct intensity profile and
gradient orientation pattern. These patterns can be used to
distinguish between fruit and background foliage. We present
a novel keypoint detector, called Angular Invariant Maximal
Detector, for detecting smooth round fruit such as grapes
and apples, and it has the following novel attributes which
we see as the contributions of this work:
1) The Angular Invariant Maximal is scale invariant
2) Color-agnostic, operates in most challenging situation
of green immature fruit over a cluttered green leaf
background
3) Robust to partial occlusions and
4) Detects a variety of round fruit, such as grapes and
apples with very high precision and with little need for
manual parameter specification.
The rest of the paper is organized as follows: 1) a section
on related work on fruit detection, 2) implementation details
of the Maximal Orientation Detector, 3) a description of the
data-sets and the experimental setup and 4) the results and
conclusion.
II. R ELATED W ORK
Current approaches for detecting fruit in images are based
on three different types of visual cues. The three different
cues of fruit appearance correspond to color, shape, and
texture.
Fruit detection methods based on color are useful only for
segmenting fruit that are of different color to the green leaf
background. Examples have been shown for the following
fruit - mangoes (Payne et al. [15] ), apples and grapes (Dunn
and Martin [5]).

Shape and contour-based approaches overcome some of

the limitations of color-based methods. Bansal et al. [2]
describes a method for detecting immature green citrus fruits,
by using the symmetrical shape of smooth spherical fruit.
Rabatel & Guizard [11] present an approach to detect berry
shapes from visible fruit edges using an elliptical contour
model. Sengupta and Lee [16] propose a method to detect
fruit using shape analysis. These methods give reasonable
performance in uncluttered environments where occlusions
do not distort the contour of the fruit.
Texture-based methods of fruit detection, are typically
associated with an external illumination source. This was
exploited by Grossetete et al. [6], who demonstrate that a
hand-held device can be used to measure the size of isolated
clusters of fruit. Similar approaches detect the shading on
apples (Wang et al. [18]) and oranges (Swanson et al. [17]).
This paper presents a texture-based approach for curved
fruit detection, that unlike existing work is scale invariant,
is robust to partial occlusions and requires little parameter
adjustment when switching between crops and varieties.
III. A PPROACH
Several challenges are associated with fruit detection in
unstructured field environments. They are
1) Large variation in lighting using natural illumination
causes inconsistent shading and high contrast regions
in the image.
2) Lighting variation causes low contrast between the fruit
and background vegetation.
The large variation in natural lighting can be overcome to
some extent by imaging with a well-designed flash and
camera pair. However, even with controlled lighting, there
is still the problem of low-contrast between the fruit and
background vegetation. The fruit detection mechanism described in this paper addresses these challenges and is able
to successfully count fruit in cluttered vineyard and orchard
environments.
A. Fruit Detection Overview
The two approaches for finding fruit locations in images
are: a thorough search of each point in the image or
identifying a set of points, known as keypoints, that have
a high probability of being classified as fruit. Hung et al.
[7] and Hung et al. [8] present an approach that performs
per-pixel classification using unsupervised multi-scale RGB
and IR feature learning for fruit segmentation. Though this
approach is effective for a large variety of fruit, it is not
possible to run this in real time as it is computationally
expensive to classify each and every point in an 4288x2848
image. Instead, by first detecting keypoints with an efficient
algorithm, the search space is narrowed down from a million
points to less than a few thousand points and hence the
computation required is drastically reduced. Nuske et al.
[12] presents a technique to identify potential fruit locations
by locating radially symmetric points and Nuske et al. [14]
presents a method, known as the Invariant Maximal detector,
that uses the distinct shading on the fruit caused by flash to

detect the keypoints. Both these method provide high recall

rate and can be run in less than 0.2s per image. In this
paper we present an alternate keypoint detection algorithm
that utilizes the patterns in the intensity profile and gradient
orientation images to detect fruit with higher recall and
precision than the Radial Symmetry and Invariant Maximal
detector. The goal of keypoint detection is to quickly find
potential fruit locations with a high percentage of the true
fruit centers being detected. At this stage there can be number
of false positives being detected without detriment to eventual accuracy because these false detections can be filtered
out using a feature classification approach. The classification
method uses high-dimensional features extracted around each
keypoint that describe the visual appearance at the keypoint.
Then using a classification algorithm the true-fruit keypoints
can be identified as those whose features are similar to fruit
appearance model constructed before run-time. Using the
approach detailed above, our algorithm for fruit detection
is split into the following blocks:
1) Keypoint Detection - detection of potential fruit locations
2) Extract feature descriptors for each keypoint
3) Classify keypoints as fruit / not fruit
B. Keypoint Detection
Here, we describe the Angular Invariant Maximal. The
detector locates the fruit center by utilizing the distinct
intensity profile and gradient orientation pattern of the fruit
surface.
1) Terminology: The notation used in the remainder of
this paper is as follows:
SP - Seedpoints- points corresponding to regional maxima.
xm ,ym is the x y coordinate of the image seedpoint.
r radii from the seed point.
- angle in sector.
I
- 8-bit single channel gray-scale image.
I(xm , ym , r, ) returns the intensity value at pixel
location (xm + r cos(), xm + r sin()).
Sj - sector corresponding to scan-angle .
l(i) - single scan line in sector S.
2) Angular Invariant Maximal: The effect of specular
reflection on the smooth round shaped fruit surface is that
the fruit center has the maximum intensity and the intensity
decreases monotonically towards the edges of the fruit.
The intensity pattern formed on the fruit surface, due to
this, is visible as concentric bands of decreasing intensity in the grayscale image I and the gradient orientation
I I
, x ) at each point on the fruit surface is directed
arctan( y
towards the fruit center forming an easily identifiable angular
pattern (Figure 1(b-d), 2).
3) Keypoint Detection Steps: Fruit centers surrounded by
a consistent pattern of intensity and gradient orientation are
identified using the following steps:
To find possible fruit centers, we first locate seed points
SP = (xm , ym ) across the entire image that are the local

-45

45

-90
90

-135

135
-180

180

Fig. 2: Synthetic images of round fruit: The maximal intensity, Imax, formed
at the center of the fruit surface - is seen as a bright - white spot (a). From
this maximal intensity point the intensity drops gradually away from the
center towards the edges of the fruit. This results in (b) intensity bands of
decreasing strength (each intensity band is depicted in a different color)
and in (c) a distinct orientation pattern which varies from [180, 180],
represented using gradual variation in color tone, across the surface of the
fruit and which can be segmented into angular segments S1 to S8 .

,ym ,r,)
maxima I(xmr
|(r=0) = 0. We then examine the
region I(xm , ym , r, ) around these seed points to determine
whether they match the ideal patterns (Figure 2) found on
the fruit surface. For this, the region around each seed point
is divided into 8 sectors {S1 , . . . , S8 }. Each sector is grown
radially outward along three scan lines {l0 (r, ), l1 (r, +

12 ), l2 (r, + 6 )}. The scan lines are grown only if

the intensity of the points on the scan line decreases
monotonically with r.
the difference in the intensity of the seed point and
the points on the scan line is greater than a threshold
. The conditions are expressed as I(xm , ym , r, ) >
I((xm , ym , 0, )
While growing the scan lines, a record of the number of
intensity bands per scan line and the number of points in each
sector that have the same gradient orientation as the dominant
angle of the sector is maintained. Each band represents a
discrete drop in intensity, which we fix at 10%. The sector
is grown until one of the following conditions are met
a majority of its scan lines have already recorded at
least a certain number of bands of intensity drop
a majority of its scan lines have stopped growing
We consider this a valid center if half of the sectors have a
minimum number of intensity bands and the sectors have at
least half of their gradient orientation matching the dominant
angle of the sector.
This approach does not require any further inputs and is
effective over different grape and apple varieties of varying
sizes.

C. Feature Descriptor
Once keypoints are found we then form feature descriptors
of the visual appearance around the keypoint. Color based

descriptors that use RGB and L*a*b channels of the image

are not suitable for discriminating fruit that have a similar
color to its background. Local texture based descriptors
like- Surf (Bay et al. [3]), SIFT (Lowe [9]) or FREAK
(Alahi et al. [1]) - are invariant to lighting changes and
do not rely on fruit having distinguishable color from the
background. However, we have found the accuracy of the
commonly used SIFT (HOG) and SURF features is restricted
because they often over-emphasize the magnitude of gradient
intensity which is almost negligible on the surface of the
fruit, where the change of intensity is gradual and intensity
contrast is very low. We overcome these restrictions by using
a feature descriptor that uses the oriented gradients without
the gradient magnitude to describe the texture of the fruit.
The feature descriptor - Radial Histogram of Oriented
Gradients plus Pairwise Intensity Comparisons descriptor describes the region around the center of the fruit using
a modified histogram of oriented gradient and a pairwise
comparison of intensity bands.
1) Radial Histogram of Oriented Gradients (RadHOG):
To create the modified histogram of oriented gradients,
we extract a circular patch, for each keypoint, from the
orientation image. This patch is then subdivided into 16
angular sectors. For each sector, a histogram of 8 orientation
bins of 45 each is formed to cover the 180 to 180
range of orientations. The number of angular sectors and
orientation bins have been empirically chosen. This 16x8
gradient histogram forms the first part of the descriptor (128
dimensions).
2) Radial Pairwise Intensity Comparisons ( RadPIC ) :
To create the second part of the descriptor we utilize the
intensity pattern formed on the surface of the fruit. The
intensity value is at its highest at the center of the fruit
and it gradually decreases as the radius from the center is
increased. The intensity value of all the points at a given radii
is the constant. To create the second part of the descriptor
we extract a circular patch of radius R, for each keypoint,
from the grayscale image. On which, we construct a set
of concentric circles of increasing radii from the keypoint.
For each circle, we sum the difference between the intensity
value of the keypoint and each pixel that forms the circle.
The resulting R (optimal R for our grape datasets is 15 and
35 for our apple datasets) length string is concatenated with
the 16x8 gradient histogram to form the feature descriptor.
D. Classification of keypoint features
For classification we use the random forest classifier described in Breiman [4]. We have used random forest classifier
for its relative ease in tuning. We use the feature descriptors
computed at keypoints for a subset of images known as the
training set to build the random forest classifier. The images
for the training set are sampled randomly from a dataset and
the fruit centers in these images are manually defined. The
features corresponding to actual fruit centers serve as positive
samples while the features of keypoints that do not align with
the manually defined fruit centers are treated as negative
samples of fruit appearance for constructing the random

(a) Apples

(b) Grapes Varieties

Fig. 3: Images showing the detected keypoints for different varieties of
apples (a) and grapes (b) using the Angular Invariant Maximal (AIM)
detector. From these images we can see that the AIM detector is able to
identify fruits of different sizes and is robust to partial occlusions.

forest classifier. The generated random forest classifier is

then used to classify the candidate features extracted in each
image in the dataset as fruit or not.
IV. R ESULTS
A. Sensor Equipment and Datasets
Our imaging system consisting of a pair of RGB stereo
cameras and a pair of flashes (Figure 1) is setup to optimize low motion blur, capture increased depth-of-focus, and
uses low illumination power for fast-recycle times permitting high-frame rates. This camera and illumination design
maintains high image quality at high vehicle velocities and
enables deployment on large scales. The imaging system is
mounted onto the side of the farm vehicle facing the fruit
wall. Depending on the size of the fruit zone a distance of 0.9
to 1.5m is maintained between the imaging system and the
fruit zone. The farm vehicle was driven at 1.5m/s through
each row and the images were captured at 5Hz. Our datasets
consist of images of different wine-grape varieties- Petite
Syrah, Pinot Noir and Merlot, table grape variety- Scarlet
Royal- and apple varieties- Granny Smith, Red Delicious
and HoneyCrisp (Table I). The grape datasets were collected
from different vineyards across California, while the apple
datasets were collected from orchards in Washington State
and Pennsylvania.
B. Algorithm setup for performance evaluation
For evaluating the performance of our detection algorithm
we perform leave one out cross validation on the images in
the training set. The training set is also used, as described

Dataset

Location

Fruit
Attributes

Camera

Image
Resolution

Flash

Days to
harvest

GrannySmith
RedDelicious
HoneyCrisp

Rock Island,
WA
Rock Island,
WA
Bisleville, PA

Green

Nikon D300s

1072x712

Red

Nikon D300s

1072x712

Green

Nikon D300s

1072x712

ScarletRoyal
Pinot-Noir

Delano, CA

Green

4288x2848

90

Galt, CA

Green

4288x2848

Xenon flashlamp (5-10J)

90

Merlot

Paso Robles,
CA
Galt, CA

Green

Pointgrey
Grasshopper
Pointgrey
Grasshopper
Pointgrey
Grasshopper
Prosilica GE4000

AlienBees ABR800
ringflash
AlienBees ABR800
ringflash
AlienBees ABR800
ringflash
Xenon flashlamp (5-10J)

4288x2848

Xenon flashlamp (5-10J)

90

2800x2200

Einstein 640 mono

flashlights

90

Petite-Sirah

Green

14
60

TABLE I: Dataset details- fruit variety and fruit attributes, sensor details and field conditions.

earlier, for building the random forest feature classifier. Here

we use the training set to also evaluate accuracy. Each
training set approximately contains N=10 images that have
been randomly selected from a dataset and the fruit centers
in these images are manually labeled to serve as ground
truth for validation. The keypoints identified as fruit by our
algorithm, are true positives (TP) if they are near the groundtruth and false positives (FP) if they are not near the manually
defined fruit centers. For leave one out cross validation, we
separate an image, defined as the validation image, from
the training set, compute its keypoints and features and then
classify them using the random forest classifier that has been
built using the remaining (N-1=9) images.
C. Keypoint Detection Performance
In Table II we compare the average precision and recall
values obtained for the orientation detector with those of
Radial Symmetry and the Invariant Maximal detector for
each of the datasets. For the keypoint detector the main
goal is to identify as many fruit centers as possible while
narrowing down the search space from a million or so pixels
to few thousand (Figure 3). Therefore, the keypoint detector
should have a high recall rate. But it is also important that the
detector have a reasonable precision rate (minimum 0.05), as
a very low precision rate will result in the classifier having
to deal with an extremely imbalanced data set with the nonfruit class forming an overwhelming majority which causes
the classifier to under perform. With this in mind we note that
the Angular Invariant Maximal detector not only provides a
high recall rate (>0.95) for different grape and apple varieties
but it also provides a higher precision rate (>0.1) than the
other two methods (Table II).
The advantages of having a higher precision is clearly
seen in the classification results shown in (Figure 4). We
note that even though the recall rate of the Angular Invariant
Maximal detector is lower than the Invariant Maximal it out
performs the Invariant Maximal in the classification step with
F1 scores (>0.82) for grapes and (>0.8) for apples due to
its higher precision rate.
V. CONCLUSIONS
In this paper we have presented a novel keypoint dectector
- Angular Invariant Maximal (AIM) for detection of smooth

round fruit. The AIM keypoint detector utilizes the distinct

intensity and gradient orientation pattern formed on the
surface of the fruit. We have demonstrated that AIM works
for different sizes and different varieties of grapes and apples,
it is robust to partial occlusions and it has a higher precision
that our previous method. The increase in precision in the
keypoint detection step has boosted the overall performance
of our fruit detection system to 0.82 F1 score in grapes and
0.8 F1 score for apples. In the future we will investigate other
imaging configurations and large pools of data to further
increase accuracy of this fruit detection and classification
system.
VI. ACKNOWLEDGEMENTS
This work is supported in part by the US Department of
Agriculture under grant number 20126702119958 and the
National Grape and Wine Initiative. The authors would like
to thank the following people who assisted in collecting the
data for this paper; Qi Wang, Kyle Wilshusen, Luke Yoder,
Karen Lewis, Tara Baugher, Luis Sanchez, Anji Perry and
Franka Gabler.
R EFERENCES
[1] Alahi, A., Ortiz, R., & Vandergheynst, P. (2012). FREAK: Fast
retina keypoint. In IEEE Conference on Computer Vision and Pattern
Recognition. 2012: pp. 510-517.
[2] Bansal, R., Lee, W.S., Satish, S. (2013) Green citrus detection using
fast Fourier transform (FFT) leakage. Precision Agriculture, 14: pp.
5970
[3] Bay, H., Ess, A., Tuytelaars, T., Gool, L. (2008) Speeded-Up Robust
Features (SURF). Computer Vision and Image Understanding, 110:
pp. 346-359
[4] Breiman, Leo. (2001). Random forests. Machine learning 45.1: pp
5-32.
[5] Dunn, G., & Martin, S. (2004). Yield prediction from digital image
analysis: A technique with potential for vineyard assessments prior
to harvest. Australian Journal of Grape and Wine Research, 10: pp
196198.
[6] Grossetete, M., Berthoumieu, Y., Costa, J. P., Germain, C., Lavialle,
O., & Grenier, G. (2012). Early estimation of vineyard yield: Site
specific counting of berries by using a smartphone. In International
Conference of Agricultural Engineering. Abstract number C1915
[7] Hung, C., Nieto, J., Taylor, Z., Underwood, J., & Sukkarieh, S.
(2013a). Orchard fruit segmentation using multi-spectral feature learning. In Intelligent Robots and Systems (IROS), 2013 IEEE/RSJ
International Conference pp. 53145320.

Dataset

Angular Invariant
Maximal

Radial
Symmetry

Invariant
Maximal

Recall

Precision

Recall

Precision

Recall

Precision

Granny Smith
Honeycrisp
Red Delicious

0.94
0.96
0.84

0.26
0.31
0.32

0.86
0.91
0.91

0.11
0.05
0.06

0.99
0.99
0.99

0.01
0.01
0.01

Mean

0.91

0.30

0.89

0.07

0.99

0.01

Pinot-Noir
Petite-Sirah
Merlot
Scarlet-Royal

0.94
0.94
0.96
0.98

0.2
0.10
0.2
0.14

0.89
0.92
0.91
0.84

0.14
0.06
0.09
0.16

0.94
0.98
0.84
0.97

0.12
0.08
0.09
0.09

Mean

0.96

0.16

0.89

0.11

0.93

0.01

TABLE II: Comparison of keypoint detection performance of the Angular Invariant Maximal, Radial Symmetry and Maximal Detector.

Classification Performance

F1 Score

0.9
0.8
0.7
0.6
Invariant Maximal
Angular Invariant Maximal
Radial Symmetry

0.5
0.4

Granny Smith Red Delicious

Honeycrisp

Scarlett-Royal

Dataset

Merlot

Petite-Sirah

Pinot-Noir

Fig. 4: Comparison of the overall classification performance- F1 score- for each detector (Angular Invariant Maximal, Radial Symmetry and Maximal
Detector).

[8] Hung, C., Underwood, J., Nieto, J., & Sukkarieh, S. (2013b). A feature
learning based approach for automated fruit yield estimation. In9th
International Conference on Field and Service Robotics (FSR).
[9] Lowe, D. G. (2004). Distinctive image features from scale-invariant
keypoints. International Journal of Computer Vision, 60: pp. 91110.
[10] Loy, G., & Zelinsky, A. (2003). Fast radial symmetry for detecting
points of interest. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 25: pp. 959973.
[11] Rabatel, G, and C Guizard. (2007). Grape berry calibration by computer vision using elliptical model fitting. European Conference on
Precision Agriculture, 6: pp. 581- 587.
[12] Nuske, S., Achar, S., Bates, T., Narasimhan, S., & Singh, S. (2011).
Yield estimation in vineyards by visual grape detection. InProceedings
of the 2011 IEEE/RSJ International Conference on Intelligent Robots
and Systems. pp. 2352-2358
[13] Nuske, S., Gupta, K., Narasimhan, S., & Singh, S. (2012). Modeling
and calibration visual yield estimates in vineyards. In Proceedings of

the International Conference on Field and Service Robotics.

[14] Nuske, S.; Wilshusen, K.; Achar, S.; Yoder, L.; Narasimhan, S.; Singh,
S. (2014). Automated Visual Yield Estimation in Vineyards. Journal
of Field Robotics, 31: pp. 837860
[15] Payne, A., Walsh, K., Subedi, P., Jarvis, D. (2013) Estimation of
mango crop yield using image analysis-segmentation method. Computers and Electronics in Agriculture, 91: pp. 57-64
[16] Sengupta S., Lee W.S. (2012) Identification and Determination of
the Number of Green Citrus Fruit under Different Ambient Light
Conditions. Presented at International Conference of Agricultural
Engineering CIGR
[17] Swanson, M., Dima, C., & Stentz, A. (2010). A multimodal system
for yield prediction in citrus trees. In ASABE Annual International
Meeting, Pittsburgh, PA.
[18] Wang, Q., Nuske, S., Bergerman, M., & Singh, S. (2012). Automated
crop yield estimation for apple orchards. In Proceedings of the
International Symposium on Experimental Robotics.