IEEE Robotics & Automation Magazine - December 2023 - 63

pyramid structures of PANet [12] that aim to strengthen the
features extracted from the backbone and further improve
the detection capability. The head predicts targets of different
sizes based on the obtained feature maps.
The following version, YOLO v6, introduced the EfficientRep
Backbone and Rep-PAN Neck by replacing the CSPBlock
used previously with RepBlock [4]. The new structure
has a decoupled head, which adds layers separating the features
from the final head and has been found able to increase
performance. It uses an anchor-free paradigm and SimOTA,
a simplified version of optimal transport assignment [6], to
enhance the training speed and detection accuracy.
In the YOLO v7 version, several changes and enhancements
were made. Notably, extended efficient layer aggregation
networks (E-ELANs) [22] serve as the backbone
to improve the learning ability. Compared to the original
ELAN architecture, the E-ELAN changes only the architecture
in the computational block, as the architecture of the
transition layer is unchanged. The authors of [22] proposed
a new compound scaling method to employ parameters and
computation more efficiently and planned a reparameterized
model that can optimize the scaling process and be applied
to concatenation-based models. The head consists of an auxiliary,
a lead head, and a soft label assigner for coarse and
fine labels.
The efficacy of the trained models was evaluated in terms
of three criteria: 1) localize accurately the stem in an image,
2) perform classification of the two states dry and wet, and
3) have a stable transition from dry to wet in a video setting.
NETWORK TRAINING PARAMETERS
Since the three YOLO networks employ distinctive structures,
we trained each version from scratch on 7,759 training
images in two ways: baseline and tuned. The baseline variation
features the baseline hyperparameters provided by the
developers of the methods in their respective repositories. In
the tuned variation, we performed tuning of selected hyperparameters
to assess the impact of different data augmentations
on each model's performance and training time. The
tuned hyperparameters include a decrease in saturation augmentations
(.03.
augmentations (),45!
c
probability), an introduction of rotation
and a deletion of the mix-up and
paste-in augmentations. The baseline hyperparameters feature
more emphasis on color space augmentations and less on
spatial-level transformations, whereas our tuned hyperparameters
place more emphasis on spatial-level transformations
and less on color space augmentations. We justify our tuning
choices by reiterating the findings of [18], where color is not
an invariant feature in wetness detection. By training a model
on more spatial-level and fewer color augmentations, we force
the model to be more sensitive to color space discriminant
features and less sensitive to shapes since every sample is
unique. All models were trained on a Tesla P100 GPU for 80
epochs, with an initial learning rate of 0.01. For optimization,
a stochastic gradient descent (SGD) optimizer was applied,
with a momentum of 0.937 and a weight decay of 0.0005.
EVALUATION OF NETWORKS' INFERENCE CAPABILITIES
We evaluated each trained network on the basis of test classification
accuracy, mean average precision (mAP), and inference
speed. Inference was performed on both cloud and edge
computing hardware with test images and on edge computing
hardware with a test video. Table 4 presents the training
results of all the YOLO models, with their baseline and tuned
hyperparameter variants. Both the baseline and tuned YOLO
v5 models achieved the highest mAP, at 0.5:0.95, within their
corresponding variants. This means that both YOLO v5 models
predicted accurately and with high confidence (. )095.
in
their detection. Out of all the models trained, the tuned
YOLO v7 scored the highest mAP, 0.5, indicating its ability
to detect stems close to their ground truth bounding boxes.
Regarding the models' cloud-based inference, we can
observe from Table 5 that our tuning of hyperparameters and
augmentations increased the classification performances of
YOLO v5 and YOLO v7. Additionally, the inference speed
increases marginally for both models. Conversely, YOLO v6
degraded in both classification performance and inference
speed under our tuning parameters. Our results for cloudbased
inference using a Tesla P100 GPU yielded inference
speeds within approximately 30 ms/frame (30fps)
$
and
thus prove that real-time water detection with our approach
is possible given equivalent or more powerful computing
hardware.
With respect to edge-based inference, and with reference
to Table 6, both YOLO v5 models are able to perform predictions
within every half second on the Nvidia Jetson Xavier NX,
and both models also attained a 100% classification accuracy
metric on the test images. We observe that all the other models
TABLE 4. The three distinctive YOLO networks and variations considered in this work and the training results on our
developed stem image dataset with base and tuned hyperparameters and augmentations.
NETWORK
YOLO v5 base
YOLO v6 base
YOLO v7 base
YOLO v5 tuned
YOLO v6 tuned
YOLO v7 tuned
SIZE (MB)
13.8
36.3
71.4
13.8
36.3
71.4
TRAINING TIME (H)
2.1
6.1
10
2.4
6.6
10
EPOCHS
80
80
80
80
80
80
RECALL
0.999
0.94
0.999
0.999
0.91
0.999
PRECISION
1
0.995
1
0.995
0.995
0.994
MAP AT 0.5 MAP AT 0.5:0.95
0.995
0.995
0.998
0.995
0.995
0.998
0.929
0.91
0.91
0.885
0.881
0.884
DECEMBER 2023 IEEE ROBOTICS & AUTOMATION MAGAZINE
63

IEEE Robotics & Automation Magazine - December 2023

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2023