E alua ing he Ene gy E iciency o Few-Sho Lea ning o Objec
De ec ion in Indus ial Se ings
Geo gios Tsoumplekas∗, Vladisla Li†, Ilias Siniosoglou∗‡, Vasileios A gy iou†, So i ios K. Goudos§,
Ioannis D. Moscholios¶, Panagio is Radoglou-G amma ikis‡∥ and Panagio is Sa igiannidis∗‡
Abs ac —In he e e -e ol ing e a o A i icial In elligence
(AI), model pe o mance has cons i u ed a key me ic d i ing
inno a ion, leading o an exponen ial g ow h in model size and
complexi y. Howe e , sus ainabili y and ene gy e iciency ha e
been c i ical equi emen s du ing deploymen in con empo a y
indus ial se ings, necessi a ing he use o da a-e icien ap-
p oaches such as ew-sho lea ning. In his pape , o alle ia e
he bu den o leng hy model aining and minimize ene gy
consump ion, a ine uning app oach o adap s anda d objec
de ec ion models o downs eam asks is examined. Subsequen ly,
a ho ough case s udy and e alua ion o he ene gy demands o
he de eloped models, applied in objec de ec ion benchma k
da ase s om ola ile indus ial en i onmen s, is p esen ed.
Speci ically, di e en ine uning s a egies, as well as u iliza ion
o ancilla y e alua ion da a du ing aining, a e examined, and
he ade-o be ween pe o mance and e iciency is highligh ed
in his low-da a egime. Finally, his pape in oduces a no el
way o quan i y his ade-o h ough a cus omized E iciency
Fac o me ic.
Index Te ms—Few-Sho Lea ning, G een AI, Deep Lea ning,
Model Op imiza ion, Objec De ec ion, Indus ial Image Da a
I. INTRODUCTION
O e he las ew yea s, he global indus ial ecosys em
has inc easingly inco po a ed AI capabili ies o ope a ional
e iciency and inno a ion, pa icula ly in a eas like p edic i e
main enance, quali y con ol, logis ics, and supply chain op-
imiza ion. Deep Lea ning (DL) models ha e p o en o be
essen ial o iden i ying ou lie s, anomalies, and i egula i ies
in indus ial en i onmen s, enabling decisi e decision-making
and p oac i e sa e y ope a ions [1]. Howe e , he inna e na-
u e o AI p esen s challenges, especially ega ding ene gy
consump ion. As complex models p ocess la ge da ase s, he
ene gy equi ed o ain and apply hem inc eases, aising en i-
onmen al, sus ainabili y, and scalabili y conce ns. The e o e,
∗G. Tsoumplekas, I. Siniosoglou and P. Sa igiannidis a e
wi h he R&D Depa men , Me aMind Inno a ions P.C., Kozani,
G eece - E-Mail: {g soumplekas, isiniosoglou,
psa igiannidis}@me amind.g
†V. Li and V. A gy iou a e wi h he Depa men o Ne wo ks and Digi al
Media, Kings on Uni e si y, Kings on upon Thames, Uni ed Kingdom -
E-Mail: { .li, asileios.a gy iou}@kings on.ac.uk
‡I. Siniosoglou, P. R. G amma ikis and P. Sa igiannidis a e wi h he Depa -
men o Elec ical and Compu e Enginee ing, Uni e si y o Wes e n Mace-
donia, Kozani, G eece - E-Mail: {isiniosoglou, p adoglou,
psa igiannidis}@uowm.g
§S. K. Goudos is wi h he Physics Depa men , A is o le
Uni e si y o Thessaloniki, Thessaloniki, G eece - E-Mail:
[email p o ec ed]
¶I. D. Moscholios is wi h he Depa men o In o ma ics and Telecom-
munica ions, Uni e si y o Peloponnese, T ipoli, G eece - E-Mail:
[email p o ec ed]
∥P. R. G amma ikis is wi h he Depa men o Resea ch and De elopmen ,
K3Y L d., So ia, 1000, Bulga ia - E-Mail: [email p o ec ed]
he ene gy e iciency o AI sys ems is a c i ical a ea o ocus,
necessi a ing esea ch o c ea e mo e ene gy-e icien and cos -
e ec i e AI models wi hou comp omising pe o mance and
accu acy, which could lead o b oade adop ion o hem in
indus ial applica ions.
Few-sho lea ning [2] (FSL) has ecen ly eme ged as a
lea ning pa adigm ha enables models o lea n om a limi ed
amoun o da a, ackling he ex ensi e esou ce demands o
s anda d AI models. Fo ypical lea ning asks wi hin he
indus ial ecosys em, such as objec de ec ion [3], da a can
be sca ce o expensi e o acqui e due o p i acy egula ions.
A he same ime, ha dwa e and bandwid h limi a ions inhe en
o edge de ices used in hese se ings ende model aining
om sc a ch p ohibi i ely ine icien .
A simple ye e ec i e FSL app oach o handle hese es ic-
ions would be o ollow a ine uning-based app oach, whe e a
p e ained model is adap ed o he new limi ed da a by e ain-
ing he model pa ially o in o al wi h hem. This app oach is
s aigh o wa d and can be e ec i ely applied in a da a-sca ce
egime. I also equi es only a ac ion o he esou ces needed
compa ed o model aining om sc a ch. As a esul , i can
po en ially se e as a p ac ical and e icien solu ion o asks
such as ew-sho objec de ec ion in se ings whe e ene gy
e iciency is c i ical. Howe e , quan i ying and enhancing he
ene gy e iciency o ine uning-based FSL models has been
sca cely explo ed in he li e a u e.
This pape explo es he balance be ween he aining pe -
o mance and ene gy e iciency o ine uning-based me hods
o ew-sho objec de ec ion in indus ial se ings. Fo his
pu pose, a ious YOLO 8 models a e ained using di e en
ine uning s a egies in h ee benchma k image da ase s om
ola ile indus ial en i onmen s, and bo h de ec ion pe o -
mance and ene gy e iciency a e sys ema ically e alua ed,
p o iding insigh s in o he ade-o be ween hese wo ob-
jec i es. Finally, a no el me ic is p oposed o be e cap u e
he a o emen ioned ade-o . The o e all con ibu ions o his
pape can be summa ized as ollows:
•In oduces a no el me ic, E iciency Fac o , o quan i y
and co ela e he ene gy consump ion s pe o mance
ade-o o FSL models.
•P esen s a ho ough e alua ion o a ious ine uned mod-
els’ pe o mance agains hei ene gy e iciency.
•C ea es an in-dep h compa a i e s udy o he e ec o
di e en benchma k da ase s on a widely used objec
de ec ion model.
•E alua es he e icacy o ine uning-based FSL as a
p ominen me hod o minimize aining ime and ene gy
43
2024 IEEE 3 d Real-Time and In elligen Edge Compu ing Wo kshop (RAGE)
979-8-3503-6334-0/24/$31.00 ©2024 IEEE
DOI 10.1109/RAGE62451.2024.00016
2024 IEEE 3 d Real-Time and In elligen Edge Compu ing Wo kshop (RAGE) | 979-8-3503-6334-0/24/$31.00 ©2024 IEEE | DOI: 10.1109/RAGE62451.2024.00016
Au ho ized licensed use limi ed o: In e na ional Hellenic Uni e si y. Downloaded on No embe 04,2025 a 15:18:25 UTC om IEEE Xplo e. Res ic ions apply.
consump ion in AI models.
The es o his pape is o ganized as ollows: he ela ed
wo k is discussed in Sec ion II, ollowed by an o e iew
o he me hodology in Sec ion III. Sec ion IV p o ides a
comp ehensi e analysis o he a ailable da a while measu ing
he ene gy e iciency and pe o mance o he models. Sec ion
Vo e s concluding ema ks.
II. RELATED WORK
A. Objec De ec ion
Objec De ec ion has made g ea p og ess h ough he
de elopmen o models such as YOLO 8, Mask RCNN, and
Fas RCNN, each making dis inc con ibu ions o he a ea.
YOLO 8, being a successo o YOLO 5, p o ides signi -
ican enhancemen s in e ms o bo h accu acy and speed
[4] Howe e , he ene gy e iciency o YOLO 8 s ill emains
a challenge, especially in edge compu ing si ua ions. Mask
RCNN is an ex ension o Fas e RCNN [5] ha allows o
pixel-le el segmen a ion [6], which means i can accu a ely
loca e objec s and pe o m ins ance segmen a ion. I pe o ms
excep ionally well in si ua ions ha demand p ecise de ec ion,
such as medical pic u e analysis. Ne e heless, he in ica e
s uc u e o he sys em esul s in inc eased compu a ional
expenses, hence a ec ing i s ene gy e iciency in se ings
wi h limi ed esou ces. Fas RCNN, an an eceden o Fas e
RCNN, in eg a es selec i e sea ch wi h a deep CNN, he eby
diminishing he duplica ion in ea u e compu ing [7]. Al hough
i ep esen ed a no able ad ancemen in he e iciency o objec
iden i ica ion, i s dependence on ex e nal egion p oposal
me hods hinde s i s speed and ene gy e iciency in compa ison
o mo e in eg a ed models such as Fas e RCNN.
B. Few-Sho Objec De ec ion
While mos FSL esea ch has adi ionally ocused on
image classi ica ion, in ecen yea s he e is an inc easing
in e es in he de elopmen o no el ew-sho objec de ec ion
(FSOD) algo i hms. One o he i s app oaches owa ds ha
di ec ion has p oposed a eweigh ing module ha ans o ms
he ex ac ed ea u e ep esen a ion and is join ly ained wi h a
YOLO de ec ion model in a wo-s ep aining p ocedu e [8]. A
wo-s ep aining app oach has also been explo ed in [9] which
demons a es he e ec i eness o ine uning in FSOD wi hou
he need o ex e nal modules. Mo e ecen ly, ecas ing he
objec de ec ion p oblem as an image classi ica ion p oblem
and lea ning new objec classes in an ad e sa ial manne has
also been p oposed [10]. Addi ionally, me a-lea ning has been
le e aged o enable lea ning ask-speci ic and ask-agnos ic
model pa ame e s in he con ex o FSOD [11]. Finally, FSOD
has also been ex ended o no el se ings ia i s combina ion
wi h inc emen al lea ning [12] and domain adap a ion.
C. AI Model Ene gy E iciency
The ield o G een and Ene gy e icien AI is s eadily
e ol ing, wi h esea ch cu en ly ocusing on mi iga ing he
ecological consequences associa ed wi h aining ex ensi e
ML models and aiming o quan ize and subsequen ly ackle he
ene gy needs o mode n AI. In pa icula , he ise in model size
and complexi y has led o a la ge inc ease in ene gy usage and
ca bon emissions [13] and consequen ly, se e al me hods ha e
been sugges ed o educe hese impac s, such as calcula ing he
ca bon emissions caused by AI models and c ea ing ools o
assess he en i onmen al impac o model aining. Fede a ed
Lea ning (FL) has also been in es iga ed as a means o mi i-
ga e ene gy usage [14]. Howe e , cons ain s in compu a ional
capaci y and he equi emen o in e -de ice communica ion
p esen obs acles. Al hough G een AI is c ucial, he e emains
a dea h o esea ch on employing FSL as a iable al e na i e.
While me a-lea ning has achie ed excellen esul s in e ms
o pe o mance in FSL, i s la ge compu a ional complexi y
ende s i inhibi ing o ene gy e icien applica ions. Howe e ,
ecen ly, ans e lea ning me hods, such as ine- uning, ha e
eme ged as iable al e na i es in his con ex , due o hei high
pe o mance and low compu a ional cos .
III. METHODOLOGY
A. Model A chi ec u e
Gi en ha ou ocus is owa ds models ha adhe e o he
p inciples o G een AI[15] and a he same ime demons a e
s ong gene aliza ion pe o mance, YOLO 8, he la es e sion
o he You-Only-Look-Once (YOLO) objec de ec ion mod-
els amily, is used. Compa ed o i s p edecesso , YOLO 5,
YOLO 8 in oduces a highly e icien ancho - ee objec de-
ec ion app oach ha leads o inc eased pe o mance.
Rega ding i s a chi ec u e, YOLO 8 consis s o a backbone
ea u e ex ac o , used o ex ac meaning ul ea u e ep esen-
a ions om he images, ollowed by he model’s head ne wo k
ha p oduces he inal p edic ions. As o he ea u e ex ac o ,
i is based on a modi ied e sion o CSPDa kne 53[16]. I s
s uc u e ollows ha o ea u e py amid ne wo ks (FPNs)
[17], which enables he iden i ica ion o objec s o a ying
sizes and scales wi hin an image by ex ac ing ea u es a
mul iple scales. On he o he hand, he head ne wo k consis s
o a se ies o con olu ional laye s ollowed by h ee di e en
de ec ion modules whose inpu s a e ea u es ex ac ed om
di e en le els o he FPN, allowing o mul i-scaled objec
de ec ion.
B. Few-sho Objec De ec ion
The main objec i e o FSOD is aining o models ha
a e capable o quickly adap ing o no el asks gi en only a
minimal numbe o aining samples wi hin each new ask.
In he con ex o FSOD, a s anda d app oach is o conside
wo di e en se s o da a wi h di e en objec classes, he
base classes and he no el classes. In his case, i is ealis ic
o assume ha he examined objec de ec ion model has
been ained on a aining se ha consis s o abundan da a
belonging o hese base classes. Mo e speci ically, he aining
se can be deno ed as:
D ain ={(xi, yi)}|D ain|
i=1 (1)
44
Au ho ized licensed use limi ed o: In e na ional Hellenic Uni e si y. Downloaded on No embe 04,2025 a 15:18:25 UTC om IEEE Xplo e. Res ic ions apply.
𝑓θ
YOLO 8
Base Da ase
Objec
de ec ion on
base classes 𝑓θ'
YOLO 8
que y-se
suppo -se
No el Task
S age 1: P e aining on base classes S age 2: Fine uning on no el classes
Objec
de ec ion on
no el classes
Fig. 1: Illus a ion o he wo-s ep aining p ocedu e based on base class
p e aining and no el class ine uning
whe e xi∈RMis he i- h aining image, yi∈ {0,1, ..., NB−
1}is i s label, and NBis he o al numbe o base classes.
Consequen ly, a ained model θis p oduced.
Following aining o he model on D ain, he nex s ep
is i s adap a ion o he no el classes. In gene al, hese no el
classes can be o mula ed as pa o a ew-sho ask ha
consis s o only a small o numbe o no el class images
a ailable du ing adap a ion and an a bi a y numbe o no el
class images used o e alua ion o he model. Speci ically,
o a gi en ask τwi h NNno el classes, i can be spli in o
a suppo se S={(xi, yi)}|S|
i=1 used o adap a ion, and a
que y se Q={(xj, yj)}|Q|
j=1 used o model e alua ion in
his ask. Adhe ing o he s anda d me hodology o FSL, N-
way K-sho asks a e examined, which consis o Nno el
classes and he e a e K aining samples o each no el class
in he ask’s suppo se (as a esul |S|=NK).
To adap o ask τ ha con ains he no el class samples,
he ained model θis u he ained on τ’s suppo se S,
p oducing an adap ed model θ′which is hen e alua ed on
τ’s que y se Q. Finally, he model’s o e all pe o mance is
i s mean pe o mance ac oss all asks {τi}T
i=1, whe e Tis he
o al numbe o asks, epo ed along wi h he co esponding
s anda d de ia ion.
C. Few-Sho Lea ning ia Model Fine uning
Based on he a o emen ioned s anda d me hodology o
FSOD, one ques ion ha ypically a ises in hese se ings is
how o e ec i ely and e icien ly ob ain θ′ om θusing
S. One o he simples app oaches in his case, which has
ecen ly achie ed compe i i e esul s compa ed o mo e com-
plex solu ions in a ious se ings, is o simply ine une he
ained model θin S. O e all, his ine uning app oach can
be seen as a wo-s ep p ocedu e. In he i s s ep, he model
is ained on D ain, emula ing a o m o model p e aining
and p oducing θ. In he second s ep, he p e ained model is
hen ine uned on each ask’s suppo se S, p oducing he inal
ine uned model θ′. I is also wo h men ioning ha a new
ou pu laye is in oduced o each ask due o he di e en
classes included in each o hem. Figu e 1also illus a es his
wo-s ep app oach.
A common conside a ion when ine uning is used o adap
a model in he con ex o FSL is deciding on a ull s pa ial
ine uning app oach. To u he examine hese wo di e en
app oaches bo h in e ms o downs eam ask pe o mance
as well as compu a ional e iciency du ing aining, we em-
ploy h ee di e en model ine uning a ia ions, le e aging
YOLO 8’s in e nal s uc u e: (a) ull ine uning o he whole
model, including bo h backbone and head, (b) pa ial ine un-
ing including he model’s head only, and (c) pa ial ine uning
including he model’s de ec ion modules only.
IV. EXPERIMENTAL RESULTS
A. Expe imen al Con igu a ion
To allow o a ai compa ison ac oss he h ee a o emen-
ioned ine uning app oaches, model p e aining is ixed in
all cases and a p e ained e sion o YOLO 8 ained on
he MSCOCO da ase is used, speci ically YOLO 8n which
consis s o 3.2M pa ame e s. In he case o ull model ine un-
ing all 3.2M pa ame e s a e adap ed, while de ec ion modules
ine uning in ol es he adap a ion o ≈750K pa ame e s, and
head ine uning in ol es he adap a ion o ≈1.7M pa ame e s.
In ou expe imen s, ull ine uned models a e deno ed as ull,
models wi h ine uned heads a e deno ed wi h head, and
models wi h ine uned de ec ion modules a e deno ed wi h de .
Addi ionally, du ing model ine uning, objec classes ha e
only Ksuppo se samples, wi h K∈ {1,2,3,5,10,30}. The
numbe o ine uning epochs is also adap ed based on K. Fo
K= 1, he numbe o epochs is 10, o K∈ {2,3,5}i is 30,
and inally o K∈ {10,30}i is 200. The use o a alida ion
se o measu e model pe o mance a e each ine uning epoch
is also examined (despi e po en ial compu a ional o e head)
o acili a e selec ing he bes -pe o ming model o be used
du ing model es ing. In ou expe imen s, models using a
alida ion se a e deno ed as bes , while he es a e deno ed as
las . To ensu e obus compa isons, each model is e alua ed
in h ee di e en downs eam asks, wi h he mean alue and
s anda d de ia ion epo ed o each me ic ac oss hese asks.
Finally, he op imized AI models a e es ed on an edge
ecosys em, conside ing esou ce limi a ions. A mid- ange
esou ce-cons ained edge de ice wi h a 12 h Gen In el i7
CPU, 16GB memo y, in eg a ed g aphics, and Ubun u 22.04
was used.
B. Da ase s
Fo he ollowing expe imen s, h ee di e en da ase s
ela ed o objec de ec ion o objec s commonly ound in
indus ial se ings we e u ilized o ine une he objec de ec ion
models, in an e o o es ablish ecogni ion and p oac i eness
in sa e y upkeeping and obus decision-making:
•Pe sonal P o ec i e Equipmen De ec ion [1]: Aims o
help AI de ec o s loca e and iden i y a ious PPE used
by i s esponde s o enhanced sa e y.
•Cons uc ion Sa e y De ec ion [18]: T ains AI mod-
els o iden i y PPE p esence o absence in indus-
ial/cons uc ion se ings.
•Fi e De ec ion: Focuses on aining AI o loca e, ecog-
nize, and classi y i es o p oac i e sa e y measu es
The abo emen ioned da ase s we e u ilized o ine une he
le e aged AI models o his wo k in he p emise o e alua ing
45
Au ho ized licensed use limi ed o: In e na ional Hellenic Uni e si y. Downloaded on No embe 04,2025 a 15:18:25 UTC om IEEE Xplo e. Res ic ions apply.
Fig. 2: Da ase P e iew. (a) PPE, (b) Cons uc ion Sa e y, (c) Fi e De ec ion
he inal models’ pe o mance s ene gy e iciency o objec
de ec ion. The a ibu es o hese da ase s a e summa ized in
Table I, while Figu e 2shows some examples o hese da ase s’
images.
C. E alua ion Me ics
Fo he e alua ion o he examined models, bo h pe o -
mance and e iciency should be aken in o conside a ion. In
his se ing, model pe o mance e e s o he ypical e alua ion
me ics used o measu e he gene aliza ion capabili ies o a
model on a gi en da ase and, consequen ly, can be quan i ied
using Mean A e age P ecision (mAP):
mAP =1
N
N
X
i=1
APi(2)
whe e APiis he a e age p ecision o class i, and Nis he
o al numbe o classes in he da ase .
On he o he hand, model e iciency e e s o a measu e
o how ene gy-e icien he examined model is. Consequen ly,
ene gy consump ion, measu ed in wa -hou s (Wh), is cal-
cula ed du ing he model’s ine uning s ep, using Py hon’s
CodeCa bon lib a y.
Finally, o accoun o bo h model pe o mance and e -
iciency, a no el me ic, E iciency Fac o (EF ), is in o-
duced ha akes in o conside a ion bo h mAP and ene gy
consump ion. Speci ically, op imal models should demons a e
high pe o mance (high mAP) as well as high e iciency (low
ene gy consump ion). As a esul , EF can be o mula ed as:
EF =mAP
1 + EC (3)
whe e mAP ∈[0,100] is in i s pe cen age o m, and EC ∈
(0,+∞)is he model’s ene gy consump ion measu ed in W h.
A cons an alue is also included in he denomina o so ha
EF is bounded in [0,100), wi h highe EF alues gi en o
models ha achie e bo h high mAP alues du ing es ing and
low ene gy consump ion du ing aining.
D. Expe imen al Resul s
Model Pe o mance. Table II shows he pe o mance o
all six di e en ine uning combina ions in e ms o mAP o
he h ee examined da ase s. I is e iden ha in almos all
cases, using a alida ion se du ing ine uning leads o con-
sis en ly be e -pe o ming models, ou lining he impo ance
o ca e ully selec ing he numbe o ine uning epochs. A
he same ime, inc easing he numbe o sho s also leads o
inc eased model pe o mance, which is easonable conside ing
ha he e a e mo e a ailable aining da a in he ask’s suppo
se . Finally, model pe o mance seems o be less sensi i e in
he selec ion o he ine uning s a egy wi h all h ee s a egies
leading o compa able esul s o he same numbe o sho s.
This is also illus a ed in Figu e 3, whe e he pe o mance o
models using e alua ion se s du ing ine uning can be seen o
di e en ine uning s a egies and di e en numbe s o sho s.
1 2 3 5 10 30
# Sho s
0
10
20
30
40
mAP(%)
Cons uc ion Sa e y
De ec o
Head
Full
(a) CS Da ase
1 2 3 5 10 30
# Sho s
0
1
2
3
4
mAP(%)
Fi e
De ec o
Head
Full
(b) Fi e Da ase
1 2 3 5 10 30
# Sho s
0
10
20
30
mAP(%)
PPE
De ec o
Head
Full
(c) PPE Da ase
Fig. 3: Mean mAP and s anda d de ia ion o bes models o a ying
ine uning s a egies and numbe s o sho s.
Model E iciency. Table III shows he ene gy consump ion
o he e alua ed models du ing ine uning, clea ly demon-
s a ing ha ine uning only he de ec ion modules wi hou
using a alida ion se consis en ly ou pe o ms he es o he
models in all h ee da ase s and o each numbe o sho s.
O e all, models ha do no use a alida ion se demons a e
be e ene gy e iciency compa ed o he co esponding ones
ha use, which is logical gi en ha ine uning becomes less
compu a ionally expensi e wi hou he use o he alida ion se
in each ine uning epoch. Addi ionally, educing he numbe
o model pa ame e s ha a e ine uned also leads o educed
ene gy consump ion since he ine uning p ocedu e becomes
less compu a ionally hea y.
E ec o ine uning. Figu e 4illus a es model pe o -
mance wi h espec o he ene gy consumed du ing aining
o each o he examined da ase s. O e all, op imal models in
e ms o pe o mance and ene gy e iciency should achie e
high mAP combined wi h low ene gy consump ion. As a
esul , models ha lie on he uppe le a ea o each sca e
plo a e conside ed bes . Howe e , he e is a clea ade-o
be ween pe o mance and e iciency since be e -pe o ming
models consume mo e ene gy du ing ine uning. Finally, i is
46
Au ho ized licensed use limi ed o: In e na ional Hellenic Uni e si y. Downloaded on No embe 04,2025 a 15:18:25 UTC om IEEE Xplo e. Res ic ions apply.
TABLE I: Main cha ac e is ics o he examined indus ial objec de ec ion da ase s.
Da ase Name To al Samples Classes T ain Tes Valida ion
PPE De ec ion Da ase 342 Helme , Glo es, Mask, Clo h Localize and iden i y PPE o i s esponde s 280 31 31
Wo ke -Sa e y Compu e Vision P ojec 3200 Helme s, Ves s, o he PPE Iden i y p esence/absence o PPE in indus ial se ings 2991 90 119
Fi e De ec ion Da ase 3677 Fi e scenes Loca e, ecognize, and classi y i es 3527 - 150
TABLE II: Few-sho de ec ion pe o mance (mean mAP and s anda d de ia-
ion) on each o he h ee examined da ase s.
Da ase Model Sho s
1 2 3 5 10 30
PPE
y8-de -las 9.68±1.43 6.79±2.35 10.49±1.94 10.93±1.10 16.50±1.89 19.11±1.12
y8-de -bes 9.71±1.39 10.22±1.51 12.47±1.41 13.10±0.64 16.21±0.72 21.29±0.94
y8-head-las 9.97±2.04 6.84±2.32 10.95±1.66 10.91±1.88 19.63±4.09 29.88±1.49
y8-head-bes 10.01±1.99 9.66±1.14 11.53±1.88 12.46±0.47 20.92±2.05 29.55±1.12
y8- ull-las 9.69±1.07 8.84±2.20 12.76±3.06 10.07±3.78 13.09±2.07 29.87±1.52
y8- ull-bes 9.69±1.07 9.91±2.26 12.79±2.71 12.59±1.60 14.68±2.58 31.96±1.59
Fi e
y8-de -las 1.63±0.24 1.99±0.49 1.75±0.15 1.74±0.42 1.54±0.49 1.63±0.53
y8-de -bes 1.81±0.03 1.94±0.07 1.93±0.17 2.09±0.35 2.42±0.26 3.15±0.42
y8-head-las 1.69±0.20 2.04±0.27 1.95±0.27 1.73±0.36 0.79±0.34 1.12±0.76
y8-head-bes 1.82±0.05 2.11±0.26 1.77±0.07 2.18±0.32 2.82±0.55 3.62±0.69
y8- ull-las 1.61±0.15 1.39±0.27 1.75±0.98 0.86±0.38 0.73±0.40 0.84±0.69
y8- ull-bes 1.82±0.06 1.95±0.21 2.51±0.86 1.56±0.40 2.61±0.34 2.52±1.69
CS
y8-de -las 8.54±0.98 10.30±3.41 10.88±2.64 12.79±2.70 19.48±3.03 31.64±0.74
y8-de -bes 8.62±0.85 11.44±2.91 12.67±1.65 13.81±1.60 17.17±2.46 32.02±1.95
y8-head-las 8.77±1.43 11.92±0.76 9.57±3.16 11.31±1.94 26.16±0.42 36.28±1.74
y8-head-bes 8.96±1.17 13.41±1.13 12.59±2.02 12.86±1.63 23.49±4.12 36.76±1.97
y8- ull-las 8.81±1.37 11.70±0.80 10.15±2.77 13.00±2.57 24.76±5.39 34.54±3.91
y8- ull-bes 8.92±1.23 13.30±1.22 13.09±1.89 13.89±1.66 24.67±5.21 34.19±4.06
TABLE III: Ene gy consump ion (mean Wh and s anda d de ia ion) on each
o he h ee examined da ase s.
Da ase Model Sho s
12351030
PPE
y8-de -las 0.271±0.003 0.639±0.036 0.843±0.007 1.194±0.060 12.435±0.268 30.010±0.658
y8-de -bes 0.774±0.020 2.400±0.016 2.607±0.007 2.916±0.053 24.738±0.442 42.395±1.153
y8-head-las 0.282±0.002 0.720±0.032 0.994±0.001 1.464±0.066 16.821±0.359 41.395±0.715
y8-head-bes 0.779±0.015 2.280±0.045 2.668±0.017 3.070±0.047 27.647±0.185 52.831±0.691
y8- ull-las 0.360±0.004 1.095±0.068 1.517±0.014 2.196±0.096 25.169±0.290 63.614±0.919
y8- ull-bes 0.852±0.003 2.762±0.020 3.190±0.017 3.866±0.077 36.599±0.224 74.630±1.195
Fi e
y8-de -las 0.487±0.010 0.601±0.005 0.660±0.007 0.734±0.004 3.772±0.040 10.294±0.062
y8-de -bes 2.161±0.041 6.341±0.015 6.448±0.036 6.573±0.011 43.335±0.470 48.161±0.434
y8-head-las 0.545±0.009 0.741±0.012 0.783±0.009 0.897±0.013 5.709±0.021 13.889±0.110
y8-head-bes 2.316±0.069 6.734±0.020 6.775±0.032 6.809±0.013 46.277±0.760 50.452±1.361
y8- ull-las 0.532±0.012 0.782±0.014 0.852±0.009 1.018±0.013 7.235±0.018 20.388±0.504
y8- ull-bes 2.273±0.014 6.421±0.044 6.472±0.039 6.576±0.113 43.477±0.259 54.348±1.413
CS
y8-de -las 0.443±0.007 0.857±0.011 1.095±0.025 1.677±0.028 15.681±0.088 43.325±0.673
y8-de -bes 1.855±0.077 5.588±0.031 5.620±0.053 6.111±0.160 44.621±0.703 70.761±1.686
y8-head-las 0.495±0.006 1.032±0.008 1.365±0.041 2.033±0.013 20.858±0.384 57.974±0.673
y8-head-bes 1.801±0.045 5.506±0.025 5.778±0.037 6.300±0.102 49.440±0.402 85.889±0.524
y8- ull-las 0.616±0.009 1.468±0.018 1.913±0.067 3.093±0.031 33.436±0.776 88.492±0.635
y8- ull-bes 1.986±0.097 5.977±0.045 6.222±0.040 7.263±0.079 62.464±1.427 119.86±2.998
also wo h no icing ha ine uning he model’s head seems o
cons i u e a good comp omise be ween hese wo con lic ing
objec i es, especially o an inc eased numbe o sho s.
E ec o alida ion se . To illus a e how model pe o -
mance wi h espec o ene gy consump ion is a ec ed by he
use o a alida ion se du ing ine uning he ela ed esul s
om he PPE da ase a e displayed in Figu e 5. Simila o
Figu e 4, he e seems o be a clea ade-o be ween model
pe o mance and ene gy consump ion, unde lining he chal-
lenges in de eloping FSOD models ha a e bo h sus ainable
and e ec i e. Addi ionally, in e ms o mAP, he use o a
alida ion se becomes less impo an as he numbe o sho s
inc eases. Howe e , i s use leads o a signi ican inc ease in
ene gy consump ion, ende ing i less use ul in hese scena ios.
EF as a consolida ed pe o mance-e iciency measu e.
While pe o mance and aining e iciency seem o be wo
con lic ing objec i es in he con ex o FSOD, i is impo an
o be able o compa e models in a uni ied way, aking in o
conside a ion bo h o hese desi ed p ope ies. Table IV shows
he EF alues o he examined models in all h ee da ase s
o a a ying numbe o sho s. I is mani es ha no using
a alida ion se du ing ine uning leads o inc eased EF
alues due o he educed ene gy consump ion o hese models
compa ed o he co esponding ones ha use alida ion se s.
Addi ionally, while adap ing less pa ame e s du ing ine uning
0 25 50 75 100 125
Ene gy(Wh)
10
15
20
25
30
35
mAP(%)
Cons uc ion Sa e y
Head
De ec o
Full
Head
De ec o
Full
(a) CS Da ase
0 10 20 30 40 50
Ene gy(Wh)
1.5
2.0
2.5
3.0
3.5
mAP(%)
Fi e
Head
De ec o
Full
Head
De ec o
Full
(b) Fi e Da ase
0 20 40 60
Ene gy(Wh)
10
15
20
25
30
mAP(%)
PPE
Head
De ec o
Full
1-sho
2-sho
3-sho
5-sho
10-sho
30-sho
(c) PPE Da ase
Fig. 4: mAP wi h espec o ene gy consump ion o di e en ine uning
s a egies and numbe s o sho s.
0 10 20 30 40
Ene gy(Wh)
7.5
10.0
12.5
15.0
17.5
20.0
mAP(%)
De ec o Fine uning
w/ e al
w/o e al
w/ e al
w/o e al
(a) De ec ion modules ine-
uning
0 10 20 30 40 50
Ene gy(Wh)
10
15
20
25
30
mAP(%)
Head Fine uning
w/ e al
w/o e al
w/ e al
w/o e al
(b) Head ine uning
0 20 40 60
Ene gy(Wh)
10
15
20
25
30
mAP(%)
Full Fine uning
w/ e al
w/o e al
1-sho
2-sho
3-sho
5-sho
10-sho
30-sho
(c) Full ine uning
Fig. 5: Model pe o mance wi h espec o ene gy consump ion in he PPE
da ase .
sh inks ene gy consump ion, i does no necessa ily lead o
inc eased EF alues, since i migh also be accompanied
by a loss in pe o mance. Howe e , he numbe o sho s
di ec ly a ec s EF , wi h a smalle numbe o sho s leading
o inc eased EF alues. This can be a ibu ed o he dis-
p opo iona e inc ease in ene gy consump ion as he numbe
o sho s inc eases because o he co esponding escala ion o
47
Au ho ized licensed use limi ed o: In e na ional Hellenic Uni e si y. Downloaded on No embe 04,2025 a 15:18:25 UTC om IEEE Xplo e. Res ic ions apply.
TABLE IV: Model pe o mance pe uni s o ene gy e iciency (mean EF and
s anda d de ia ion) on each o he h ee examined da ase s.
Da ase Model Sho s
1 2 3 5 10 30
PPE
y8-de -las 7.618±1.141 4.151±1.475 5.696±1.077 4.977±0.439 1.230±0.150 0.617±0.044
y8-de -bes 5.484±0.828 3.008±0.453 3.456±0.385 3.347±0.208 0.630±0.026 0.491±0.032
y8-head-las 7.784±1.605 3.982±1.363 5.488±0.832 4.417±0.698 1.106±0.248 0.705±0.033
y8-head-bes 5.635±1.154 2.948±0.374 3.146±0.519 3.060±0.079 0.730±0.071 0.549±0.027
y8- ull-las 7.132±0.808 4.255±1.204 5.071±1.226 3.160±1.228 0.500±0.079 0.462±0.024
y8- ull-bes 5.235±0.575 2.637±0.605 3.052±0.647 2.590±0.348 0.390±0.067 0.423±0.027
Fi e
y8-de -las 1.096±0.159 1.241±0.301 1.052±0.085 1.002±0.244 0.326±0.101 0.145±0.047
y8-de -bes 0.573±0.003 0.265±0.010 0.260±0.023 0.276±0.047 0.055±0.006 0.064±0.008
y8-head-las 1.097±0.131 1.173±0.146 1.094±0.157 0.913±0.197 0.119±0.050 0.075±0.050
y8-head-bes 0.549±0.007 0.272±0.033 0.227±0.010 0.280±0.041 0.060±0.013 0.070±0.013
y8- ull-las 1.051±0.108 0.777±0.147 0.945±0.531 0.426±0.183 0.089±0.049 0.039±0.032
y8- ull-bes 0.556±0.019 0.263±0.028 0.336±0.114 0.206±0.054 0.059±0.008 0.046±0.030
CS
y8-de -las 5.762±0.801 6.936±0.637 4.819±1.566 4.295±1.090 1.211±0.215 0.726±0.015
y8-de -bes 3.009±0.272 2.065±0.234 1.866±0.287 1.836±0.289 0.397±0.063 0.452±0.026
y8-head-las 5.868±0.975 5.866±0.398 4.067±1.415 3.724±0.620 1.197±0.034 0.615±0.023
y8-head-bes 3.195±0.367 2.061±0.176 1.855±0.287 1.759±0.205 0.466±0.085 0.423±0.023
y8- ull-las 5.452±0.864 4.742±0.338 3.508±1.046 3.171±0.602 0.723±0.171 0.386±0.046
y8- ull-bes 2.978±0.322 1.907±0.185 1.813±0.273 1.679±0.184 0.391±0.089 0.282±0.026
he ine uning epochs. The a o emen ioned obse a ions a e
also illus a ed in Figu e 6, whe e he EF alues o models
ine uned wi hou using a alida ion se a e displayed.
1 2 3 5 10 30
# Sho s
0
2
4
6
E iciency Fac o
Cons uc ion Sa e y
De ec o
Head
Full
(a) CS Da ase
1 2 3 5 10 30
# Sho s
0.00
0.25
0.50
0.75
1.00
1.25
1.50
E iciency Fac o
Fi e
De ec o
Head
Full
(b) Fi e Da ase
1 2 3 5 10 30
# Sho s
0
2
4
6
8
E iciency Fac o
PPE
De ec o
Head
Full
(c) PPE Fa ase
Fig. 6: Mean EF and s anda d de ia ion o las models o a ying ine uning
s a egies and numbe s o sho s.
V. CONCLUSIONS
While mode n machine and deep lea ning app oaches ha e
led o emendous s ides in a ious indus ial se ings and ap-
plica ions, hei pe o mance and ene gy e iciency e alua ion
in ealis ic ola ile and da a-sca ce indus ial en i onmen s has
been unde explo ed. In his pape , we examine how ine uning-
based FSL me hods can be le e aged o p oduce models ha
demons a e bo h high pe o mance and ene gy e iciency in
downs eam de ec ion asks wi h limi ed samples in indus ial
se ings. An empi ical s udy based on h ee di e en indus ial
da ase s is conduc ed, demons a ing he balance be ween
ene gy e iciency and he aining pe o mance in ol ed in
ine uning-based FSL app oaches while also examining how
hese a iables a e a ec ed by di e en ine uning s a egies.
Finally, a no el me ic, E iciency Fac o , is in oduced o help
quan i y he in e ac ion o model pe o mance and e iciency in
a consolida ed way. O e all, his wo k could se e as an ini ial
s ep owa ds a b oade in es iga ion o FSL’s ene gy e iciency,
which could be expanded o include s anda d echniques such
as me a-lea ning and da a augmen a ion.
ACKNOWLEDGEMENT
This p ojec has ecei ed unding om he Eu opean
Union’s Ho izon Eu ope esea ch and inno a ion p og amme
unde g an ag eemen No. 101070181 (TALON).
REFERENCES
[1] A. Sesis, I. Siniosoglou, Y. Spy idis, G. E s a hopoulos, T. Lagkas,
V. A gy iou, and P. Sa igiannidis, “A obus deep lea ning a chi ec u e
o i e igh e ppes de ec ion,” in 2022 IEEE 8 h Wo ld Fo um on In e ne
o Things (WF-IoT), 2022, pp. 1–6.
[2] G. Tsoumplekas, V. Li, V. A gy iou, A. Ly os, E. Foun oukidis, S. K.
Goudos, I. D. Moscholios, and P. Sa igiannidis, “Towa d g een and
human-like a i icial in elligence: A comple e su ey on con empo a y
ew-sho lea ning app oaches,” a Xi p ep in a Xi :2402.03017, 2024.
[3] R. Padilla, S. L. Ne o, and E. A. B. da Sil a, “A su ey on pe o -
mance me ics o objec -de ec ion algo i hms,” in 2020 In e na ional
Con e ence on Sys ems, Signals and Image P ocessing (IWSSIP), 2020,
pp. 237–242.
[4] F. M. Talaa and H. ZainEldin, “An imp o ed i e de ec ion app oach
based on YOLO- 8 o sma ci ies,” Neu al Compu ing and Applica-
ions, ol. 35, no. 28, pp. 20 939–20 954, Oc . 2023.
[5] N. Rane, “Yolo and as e -cnn objec de ec ion o sma indus y
4.0 and indus y 5.0: applica ions, challenges, and oppo uni ies,” SSRN
Elec onic Jou nal, 01 2023.
[6] K. Suja ha, K. Am u ha, and N. Vee anjaneyulu, “Enhancing objec
de ec ion wi h mask -cnn: A deep lea ning pe spec i e,” in 2023
In e na ional Con e ence on Ne wo k, Mul imedia and In o ma ion
Technology (NMITCON), 2023, pp. 1–6.
[7] T. Diwan, G. Ani udh, and J. V. Tembhu ne, “Objec de ec ion using
YOLO: challenges, a chi ec u al successo s, da ase s and applica ions,”
Mul imed. Tools Appl., ol. 82, no. 6, pp. 9243–9275, 2023.
[8] B. Kang, Z. Liu, X. Wang, F. Yu, J. Feng, and T. Da ell, “Few-
sho objec de ec ion ia ea u e eweigh ing,” in P oceedings o he
IEEE/CVF In e na ional Con e ence on Compu e Vision, 2019, pp.
8420–8429.
[9] X. Wang, T. Huang, T. Da ell, J. Gonzalez, and F. Yu, “F us a -
ingly simple ew-sho objec de ec ion. a xi 2020,” a Xi p ep in
a Xi :2003.06957.
[10] B. Li, B. Yang, C. Liu, F. Liu, R. Ji, and Q. Ye, “Beyond max-ma gin:
Class ma gin equilib ium o ew-sho objec de ec ion,” in P oceedings
o he IEEE/CVF con e ence on compu e ision and pa e n ecogni ion,
2021, pp. 7363–7372.
[11] Y.-X. Wang, D. Ramanan, and M. Hebe , “Me a-lea ning o de ec a e
objec s,” in P oceedings o he IEEE/CVF In e na ional Con e ence on
Compu e Vision, 2019, pp. 9925–9934.
[12] L. Yin, J. M. Pe ez-Rua, and K. J. Liang, “Sylph: A hype ne wo k
amewo k o inc emen al ew-sho objec de ec ion,” in P oceedings o
he IEEE/CVF Con e ence on Compu e Vision and Pa e n Recogni ion,
2022, pp. 9035–9045.
[13] P. Hende son, J. Hu, J. Romo , E. B unskill, D. Ju a sky, and J. Pineau,
“Towa ds he sys ema ic epo ing o he ene gy and ca bon oo p in s
o machine lea ning,” J. Mach. Lea n. Res., ol. 21, no. 1, jan 2020.
[14] X. Qiu, T. Pa colle , J. Fe nandez-Ma ques, P. P. de Gusmao, Y. Gao,
D. J. Beu el, T. Topal, A. Ma hu , and N. D. Lane, “A i s look in o he
ca bon oo p in o ede a ed lea ning.” J. Mach. Lea n. Res., ol. 24,
pp. 129–1, 2023.
[15] R. Schwa z, J. Dodge, N. A. Smi h, and O. E zioni, “G een ai,”
Communica ions o he ACM, ol. 63, no. 12, pp. 54–63, 2020.
[16] A. Bochko skiy, C.-Y. Wang, and H.-Y. M. Liao, “Yolo 4: Op-
imal speed and accu acy o objec de ec ion,” a Xi p ep in
a Xi :2004.10934, 2020.
[17] T.-Y. Lin, P. Doll´
a , R. Gi shick, K. He, B. Ha iha an, and S. Belongie,
“Fea u e py amid ne wo ks o objec de ec ion,” in P oceedings o he
IEEE con e ence on compu e ision and pa e n ecogni ion, 2017, pp.
2117–2125.
[18] compu e ision, “Wo ke -sa e y da ase ,” h ps://uni e se. obo low.
com/compu e - ision/wo ke -sa e y, jul 2022, isi ed on 2024-01-
23. [Online]. A ailable: h ps://uni e se. obo low.com/compu e - ision/
wo ke -sa e y
48
Au ho ized licensed use limi ed o: In e na ional Hellenic Uni e si y. Downloaded on No embe 04,2025 a 15:18:25 UTC om IEEE Xplo e. Res ic ions apply.
