IEEE Robotics & Automation Magazine - December 2023 - 64

TABLE 5. The cloud-based inference results with a Tesla P100 on the image
test set.
NETWORK
YOLO v5 base
YOLO v6 base
YOLO v7 base
YOLO v5 tuned
YOLO v6 tuned
YOLO v7 tuned
EXECUTION
TIME (S)
27.2
25
38.1
26.5
30
37.6
AVERAGE INFERENCE
SPEED PER IMAGE (MS)
10.4
25.5
21.5
8.8
30.1
21.4
CLASSIFICATION
ACCURACY
99.3%
99.5%
99.7%
100%
99.4%
100%
TABLE 6. The edge-based inference results with a Jetson Xavier NX on the
image test set.
NETWORK
YOLO v5 base
YOLO v6 base
YOLO v7 base
YOLO v5 tuned
YOLO v6 tuned
YOLO v7 tuned
EXECUTION
TIME (MIN)
8.04
19.57
38.69
7.49
17.35
46.24
AVERAGE INFERENCE
SPEED PER IMAGE (MS)
481.6
1,171.7
2,316.5
448.4
1,038.9
2,769
TABLE 7. The edge-based inference results with a
Jetson Xavier NX on the test video.
NETWORK
YOLO v5 base
YOLO v6 base
YOLO v7 base
AVERAGE INFERENCE
SPEED PER FRAME (MS)
471.4
1,020
2,863.3
YOLO v5 tuned 568.1
YOLO v6 tuned 1,030.9
2,811.6
YOLO v7 tuned
STABLE STATE
TRANSITION
No
Yes
Yes
Yes
Yes
Yes
CLASSIFICATION
ACCURACY
100%
99.5%
99.7%
100%
99.4%
100%
attained high classification metrics
as well, with the tuned YOLO v7,
in particular, scoring a 100% success
rate on the test images. Yet, given
that it is five times larger than YOLO
v5, its inference speed was about five
times slower than that of the YOLO
v5 model, making it less practical for
deployment on less powerful edge
devices. YOLO v6 experienced a slight
inference speed performance boost
with our tuning parameters.
Regarding image inference speed,
Tables 5 and 6 indicate a direct relation
between model size and inference
speed. Smaller models tend to perform
faster and are thus suited for deployment
on devices with limited computational
power. Table 7 contains the
average edge device inference speed
results for each model on a test video
taken from the dataset preprocessing
phase earlier in this work. Similar to
the image inference results, smaller
models have faster video inference
speeds. As expected, the average inference speed per frame
for each model resembles the corresponding metric in edgebased
image inference (Table 6).
QUALITATIVE ANALYSIS
Figure 12 juxtaposes the wetness detection results on a single
orange sample test case by all YOLO models trained on our
dataset. The detection confidence is approximately 0.95, and
it remains consistent with the bulk of our test results. Comparing
the baseline and tuned versions of each model, we can
observe that the bounding boxes are slightly tighter around
the stem detection for the baseline networks compared to
(a)
(b)
(c)
(d)
(e)
(f)
FIGURE 12. Correct wetness detection results on a single orange stem by all the YOLO models: (a) YOLO v5 base, (b) YOLO v5 tuned,
(c) YOLO v6 base, (d) YOLO v6 tuned, (e) YOLO v7 base, and (f) YOLO v7 tuned. The top panels correspond to the dry status, while the
bottom panels correspond to the wet status.
64 IEEE ROBOTICS & AUTOMATION MAGAZINE DECEMBER 2023

IEEE Robotics & Automation Magazine - December 2023

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