EVALUACIÓN DE ALGORITMOS DE
MACHINE LEARNING PARA CONDUCCIÓN
MACHINE LEARNING ALGORITHM EVALUATION
ON ADVANCED DRIVER ASSISTANCE
TRABAJO FIN DE GRADO
CURSO 2020-2021
BY
WENBO SUN
DIRECTOR
CARLOS GARCÍA SÁNCHEZ
GUILLERMO BOTELLA JUAN
GRADO EN INGENIERÍA INFORMÁTICA
FACULTAD DE INFORMÁTICA
UNIVERSIDAD COMPLUTENSE DE MADRID
THANKS
I am e y g a e ul o he wo eache s o hei help in his g. F om he beginning,
I didn' know and didn' know how o do i , no only ga e me he gene al di ec ion o he
p ojec , bu also ga e me some necessa y in o ma ion. In he la e pe iod, Ca los poin ed
me o he sho comings o he p ojec and encou aged me o comple e he p ojec . Fo
his I mus say ha he eache s a e e y good. I am e y g a e ul and g a e ul o hem.
A he same ime, I am e y g a e ul o he school o he help I lea ned du ing he
epidemic.
ABSTRACT
Machine lea ning algo i hm e alua ion
on ad anced d i e assis ance
In his esea ch and de elopmen p ojec , ou main pu pose is o s udy ou deep
lea ning a chi ec u es o eal- ime objec de ec ion o people and bicycles
encoun e ed in on o d i ing.
We use 4 di e en algo i hms o he same da a se , and compa e he mAPs
ob ained a e aining. And discuss which me hod is he mos accu a e, bu also conside
he ime i akes o ge wha is sui able o wha kind o scene.
The p ojec I came up wi h would like o be used in a d i ing assis ance sys em.
The sys em uses came a senso s o ge inpu , and hen uses algo i hms o assis , so ha
he sa e y o he ca is gua an eed when d i ing. A he same ime, i can un on a low-
pe o mance e sion o he machine and compa e he ps o di e en algo i hms.
Keywo ds
Compu e ision, au onomous d i ing, con olu ional neu al ne wo k, objec
de ec ion, deep lea ning, machine lea ning, ea u e selec ion and ex ac ion,
d i ing assis ance
INDEX OF CONTENTS
Chap e 1 -!In oduc ion ............................................................................................... 1!
Chap e 1.1 -!Mo i a ion ........................................................................................ 2!
Chap e 1.2 -!Objec i e ......................................................................................... 4!
Chap e 2 -!Con ibu ion o s uden ............................................................................ 5!
Chap e 2.1 -!Plan o Wo ks .................................................................................... 5!
Chap e 2.2 -!Flow Cha ........................................................................................ 6!
Chap e 3 -!Ad anced and a chi ec u e ................................................................... 7!
Chap e 3.1 -!Con olu ional Neu al Ne wo k, CNN ............................................ 7!
Chap e 3.2 -!Ta ge de ec ion ne wo k ............................................................ 10!
Chap e 3.2.1 -!Two s age ............................................................................. 10!
1.!Fas e -RCNN[4] ................................................................................... 10!
2.!Fpn[5] ................................................................................................... 13!
Chap e 3.2.2 -!One s age ............................................................................ 17!
1.!SSD[6] ................................................................................................... 17!
2.!VGG16[8] ............................................................................................ 18!
3.!MobileNe V3[9] ................................................................................... 20!
4.!YOLO 4[7] ........................................................................................... 23!
Chap e 4 -!Da a Se ................................................................................................... 30!
Chap e 4.1 -!Tshingua-Daimle [13] .................................................................... 30!
Chap e 4.2 -!INRIA Pe son Da ase [14] .............................................................. 30!
Chap e 4.3 -!C ea ed da a se .......................................................................... 30!
Chap e 5 -!Ha dwa e ................................................................................................. 31!
Chap e 5.1 -!In el(R) Co e(TM) i7-9700F CPU .................................................... 31!
Chap e 5.2 -!GEFORCE RTX 2080 Ti .................................................................... 32!
Chap e 5.3 -!Je son Nano 2GB De elope Ki [28] ........................................... 32!
Chap e 6 -!Ta ge de ec ion e alua ion ................................................................. 33!
Chap e 6.1 -!In e sec ion O e Union (IOU) ...................................................... 33!
Chap e 6.2 -!T ue Posi i e, False Posi i e, False Nega i e and T ue Nega i e
33!
Chap e 6.3 -!P ecision ......................................................................................... 34!
Chap e 6.4 -!Recall .............................................................................................. 34!
Chap e 6.5 -!AP calcula ion ............................................................................... 35!
Chap e 7 -!F amewo ks .............................................................................................. 35!
Chap e 7.1 -!Cuda ............................................................................................... 35!
Chap e 7.2 -!Py o ch ........................................................................................... 36!
Chap e 7.3 -!MMCV ............................................................................................. 37!
Chap e 8 -!Resul s o he expe imen ....................................................................... 37!
Chap e 8.1 -!Resul s o VGG16 ........................................................................... 38!
Chap e 8.2 -!Resul s o SSD MobileNe .............................................................. 40!
Chap e 8.3 -!Resul s o YOLO 4 .......................................................................... 41!
Chap e 8.4 -!Resul o FASTER RCNN+FPN ......................................................... 42!
Chap e 9 -!Compa ison o expe imen al esul s ..................................................... 43!
Chap e 9.1 -!Compa ison o esul s o high-pe o mance machines ............ 43!
Chap e 9.2 -!Compa ison o esul s o low-pe o mance machines ............. 45!
Chap e 10 -!Conclusions and u u e wo k ............................................................... 46!
Chap e 10.1 -!Conclusions .................................................................................. 46!
Chap e 10.2 -!Fu u e Wo k .................................................................................. 47!

BIBLIOGRAPHY ................................................................................................................ 49!
Appendix ........................................................................................................................ 51!
1
Chap e 1 - In oduc ion
Pedes ian De ec ion has always been a ho and di icul poin in compu e ision
esea ch. The p oblem o be sol ed by pedes ian de ec ion is o ind all pedes ians in
an image o ideo ame, including hei posi ion and size, which a e gene ally
ep esen ed by ec angula boxes, simila o ace de ec ion, which is also a ypical a ge
de ec ion p oblem.
Pedes ian de ec ion echnology has a s ong use alue. I can be combined wi h
pedes ian acking, pedes ian e-iden i ica ion and o he echnologies. I can be used
in au omo i e unmanned d i ing sys ems (ADAS), in elligen obo s, in elligen ideo
su eillance, human beha io analysis, passenge low s a is ics sys ems, and in elligence
T anspo a ion and o he ields.
Since he human body is qui e lexible, he e will be a ious pos u es and shapes,
and i s appea ance is g ea ly a ec ed by wea ing, pos u e, iewing angle, e c., and i
also aces he in luence o ac o s such as occlusion and illumina ion. This makes
pedes ian de ec ion a compu e ision A e y challenging subjec . The main p oblems
o be sol ed in pedes ian de ec ion a e:
The appea ance is e y di e en . Including iewing angle, pos u e, clo hing and
a achmen s, ligh ing, imaging dis ance, e c. Looking a he pas om di e en angles,
he appea ance o pedes ians is e y di e en . Pedes ians in di e en pos u es ha e
e y di e en appea ances. Due o he di e en clo hes ha people wea , as well as he
in luence o umb ellas, ha s, sca es, luggage and o he a achmen s, he appea ance
is e y di e en . The di e ence in ligh ing also caused some di icul ies. The human body
a a dis ance and he human body a a close dis ance a e also e y di e en in
appea ance.
2
Occlusion p oblem. In many applica ion scena ios, pedes ians a e e y dense
and he e a e se ious occlusions. We can only see a pa o he human body, which b ings
se ious challenges o he de ec ion algo i hm.
The backg ound is complica ed. Whe he indoo o ou doo , pedes ian de ec ion
gene ally aces e y complica ed backg ounds. The appea ance and shape, colo , and
ex u e o some objec s a e e y simila o human bodies, which makes he algo i hm
unable o dis inguish accu a ely.
De ec ion speed. Pedes ian de ec ion gene ally uses a complex model wi h a
la ge amoun o calcula ions. I is e y di icul o achie e eal- ime and gene ally equi es
a lo o op imiza ion.
Chap e 1.1 - Mo i a ion
One day in he u u e, new ene gy ehicles will e en ually eplace uel ehicles,
and au onomous d i ing will also eplace he d i e one day in he u u e.
The cu en le el o au onomous d i ing is o mula ed by he SAE In e na ional[27],
which is di ided in o 6 le els (L0-L5).
-Le el L0: The d i e is in ull con ol o he ehicle;
-Le el L1: The au oma ic sys em can some imes assis he d i e o comple e
ce ain d i ing asks;
-Le el L2 assis ed d i ing: The au oma ic sys em can comple e ce ain
d i ing asks, bu he d i e needs o moni o he d i ing en i onmen and
comple e he es , while ensu ing ha p oblems occu and ake o e a any
ime. A his le el, he w ong pe cep ion and judgmen o he au oma ic
sys em can be co ec ed by he d i e a any ime, and mos ca
companies can p o ide his sys em. L2 can be di ided in o di e en usage
3
scena ios based on speed and en i onmen , such as low-speed a ic jams
on he loop, as d i ing on highways, and au oma ic pa king by he d i e
in he ca ;
-Le el L3 semi-au onomous d i ing: The au oma ic sys em can no only
comple e ce ain d i ing asks, bu also moni o he d i ing en i onmen
unde ce ain condi ions, bu he d i e mus be eady o egain d i ing
con ol (when he au oma ic sys em eques s i ). The e o e, a his le el, he
d i e s ill canno sleep o ake a deep es . A e he comple ion o L2, he
esea ch ield o ca companies is ex ended om he e. Due o he
pa icula i y o L3, he mos meaning ul deploymen cu en ly seen is o
upg ade on he high-speed L2; he di e ence be ween L3 and L2 is ha
he ehicle is esponsible o pe iphe al moni o ing, and he human d i e
only needs o main ain a en ion o eme gencies.
-Le el L4 highly au oma ed d i ing: Au oma ed sys ems can comple e
d i ing asks and moni o he d i ing en i onmen unde ce ain
en i onmen s and speci ic condi ions; cu en ly, he deploymen o L4 is
mos ly based on ci y use, which can be ully au oma ed ale pa king. I
can also be di ec ly combined wi h axi se ices. A his s age, wi hin he
scope o au onomous d i ing, all asks ela ed o d i ing ha e no hing o
do wi h he d i e and passenge s. The pe cep ion o ex e nal esponsibili y
lies in he au onomous d i ing sys em, and he e a e di e en design and
deploymen ideas he e;
-Le el L5 ully au oma ed d i ing: all d i ing asks ha he au oma ed
sys em can comple e unde all condi ions.
10
Figu e[3.1.3] Fully connec ed laye ou pu esul
Chap e 3.2 - Ta ge de ec ion ne wo k
-Commonly used de ec ion can be oughly di ided in o wo ca ego ies: wo
s age and one s age. Two s age means ha he de ec ion ne wo k is di ided in o wo
s eps. The i s s ep is o ame he objec , and he second s ep is o de e mine he
classi ica ion o he objec . One s age is o di ec ly use he eg ession ne wo k o ge i s
classi ica ion and de ec ion ame based on he ex ac ed ea u es. Among hem, Fas e
R-CNN belongs o wo s age, SSD and YOLO belong o one s age.
-The ad an age o one s age is as , and he ad an age o wo s age is accu acy.
A e ge ing he de ec ion ame, i is much easie o classi y his ask han di ec ly
e u ning o he ne wo k o ge he classi ica ion.
Chap e 3.2.1 - Two s age
1. Fas e -RCNN[4]
The Fas e -RCNN model in oduces RPN (Region P oposal Ne wo k) o di ec ly
gene a e candida e egions. Fas e -RCNN can be seen as a combina ion o RPN and
Fas RCNN models, ha is, Fas e -RCNN = RPN + Fas -RCNN.

11
Fo he RPN ne wo k, a CNN model (gene ally called a ea u e ex ac o ) is
i s used o ecei e he en i e pic u e and ex ac he ea u e map. Then an N×N (3×3 in
he ex ) sliding window is used on his ea u e map, and a low-dimensional ea u e (such
as 256-d) is mapped o each sliding window posi ion. Then his ea u e is sen o wo ully
connec ed laye s, one o classi ica ion p edic ion and he o he o eg ession. Gene ally,
k a p io i boxes (ancho s, de aul bounding boxes) o di e en sizes o p opo ions a e se
o each window posi ion, which means ha k candida e egions ( egion p oposals) a e
p edic ed o each posi ion. Fo he classi ica ion laye , i s ou pu size is 2k, which means
ha each candida e a ea con ains he objec o backg ound p obabili y alue, while
he eg ession laye ou pu s 4k coo dina e alues, which indica e he posi ion o each
candida e a ea ( ela i e o each a p io i box). Fo each sliding window posi ion, hese
wo ully connec ed laye s a e sha ed. The e o e, RPN can be implemen ed using
con olu ional laye s: i s , an n×n con olu ion o ob ain low-dimensional ea u es, and
hen wo 1×1 con olu ions, which a e used o classi ica ion and eg ession espec i ely.
Figu e[3.2.1.1] CNN model (gene ally called a ea u e ex ac o )
RPN uses wo classi ica ions, only dis inguishing backg ound and objec s, bu does
no p edic he ca ego y o he objec , ha is, class-agnos ic. Since he coo dina e
alues need o be p edic ed a he same ime, du ing aining, he a p io i box mus be
ma ched wi h he g ound- u h box. The p inciple is: (1) he a p io i box wi h he highes
12
IoU o a ce ain g ound- u h box; (2) and A g ound- u h box wi h an IoU alue g ea e
han 0.7 a p io i box, as long as one is sa is ied, he a p io i box can ma ch a g ound- u h,
so ha he a p io i box is a posi i e sample (belonging o he objec ), and he g ound-
u h Fo he e u n goal. Fo hose a p io i boxes whose IoU alue wi h any g ound- u h
box is lowe han 0.3, hey a e conside ed as nega i e samples. The RPN ne wo k can be
ained sepa a ely, and he sepa a ely ained RPN model gi es many egion p oposals.
Due o he la ge numbe o a p io i boxes, many o he candida e egions p edic ed by
RPN o e lap. We mus i s pe o m NMS (non-maximum supp ession, IoU h eshold is se
o 0.7) o educe he numbe o candida e egions, and hen a ange hem in
descending o de o con idence, selec op -N egion p oposals a e used o ain he Fas
R-CNN model. The ole o RPN is o eplace he ole o Selec i e sea ch, bu i is as e , so
Fas e R-CNN can accele a e bo h aining and p edic ion.
Figu e[3.2.1.2] RPN s uc u e
13
Fas e -RCNN ollows he ollowing aining p ocess:
S ep 1: Use he p e- ained model on ImageNe o ini ialize he ea u e
ex ac ion ne wo k and ain he RPN ne wo k;
S ep 2: Use he model p e- ained on ImageNe o ini ialize he Fas -RCNN
ea u e ex ac ion ne wo k, and use he candida e box gene a ed by he ained RPN
ne wo k in s ep one as inpu o ain a Fas -RCNN ne wo k. So a , each o he wo
ne wo ks The pa ame e s o he i s laye a e no sha ed a all;
S ep 3: Use he Fas -RCNN ne wo k pa ame e s o s ep wo o ini ialize a new
RPN ne wo k, bu se he lea ning a e o he ea u e ex ac ion ne wo k pa ame e s
sha ed by RPN and Fas -RCNN o 0, e en i he unique pa ame e s o he RPN ne wo k
a e lea ned, i is ixed Fea u e ex ac ion ne wo k. A his poin , he wo ne wo ks ha e
sha ed all common con olu ional laye s;
S ep 4: Fo hose ne wo k laye s ha a e s ill sha ed, add Fas -RCNN-speci ic
ne wo k laye s, con inue aining, and ine- une Fas -RCNN-speci ic ne wo k laye s. Up o
his poin , RPN and Fas -RCNN ne wo ks comple ely sha e pa ame e s. Use Fas -RCNN o
comple e candida e ame ex ac ion and a ge de ec ion unc ions a he same ime.
2. Fpn[5]
In he pas as e cnn o a ge de ec ion, no ma e i is pn o as cnn, oi ac s
on he las laye . This is no p oblem o he de ec ion o la ge a ge s, bu he e a e some
p oblems o de ec ion o small a ge s. Because o small a ge s, when con olu ional
pooling is pe o med o he las laye , he seman ic in o ma ion is ac ually gone, because
we all know ha he me hod o mapping a oi o a ea u e map is o di ec ly di ide he
14
unde lying coo dina es by s ide, Ob iously, he la e , he smalle he map will be, and i
may e en disappea . The e o e, in o de o sol e he p oblem o mul i-scale de ec ion, a
ea u e py amid ne wo k is in oduced.
Figu e[3.2.1.3] Mul i-scale py amid
Figu e[3.2.1.3] (a) is a ai ly common mul i-scale me hod called a ea u e image
py amid. This me hod was widely used in he ea lie a i icial design ea u es (DPM), and
i was also used in CNN. I is o mul i-scale he inpu image by se ing di e en zoom a ios.
This can sol e mul iple scales, bu i is equi alen o aining mul iple models (assuming a
ixed inpu size is equi ed). E en i he inpu size is no allowed, i also inc eases he
memo y space o s o ing images o di e en scales.
Figu e[3.2.1.3] (b) is CNN. Compa ed wi h a i icially designed ea u es, cnn can
lea n mo e ad anced seman ic ea u es by i sel . A he same ime, CNN is obus o scale
changes. The e o e, as shown in he igu e, he ea u es calcula ed om he inpu o a
single scale can also be used o iden i y , Bu when encoun e ing ob ious mul i-scale
a ge de ec ion, he py amid s uc u e is s ill needed o u he imp o e he accu acy.
Judging om some o he leading me hods on he imageNe and COCO da a
se s, he ea u e image py amid me hod is used in he es , ha is, he combina ion o
Figu e[3.2.1.3] (a) and Figu e[3.2.1.3] (b). Explains ha he ad an age o each le el o
he cha ac e ized image py amid is ha i p oduces a mul i-scale ea u e ep esen a ion,
15
and he ea u es o each le el ha e s ong seman ics (because all ea u es gene a ed
by cnn), including he high- esolu ion le el (The la ges scale inpu image).
Howe e , his mode has ob ious d awbacks. Compa ed wi h he o iginal me hod,
he ime has inc eased by 4 imes, and i is di icul o use in eal- ime applica ions. Simila ly,
i also inc eases he s o age cos , which is why he image py amid is only used in he
es ing phase. Bu i i is only used in he es ing phase, hen aining and es ing will be
inconsis en in in e ence. The e o e, some ecen me hods ha e simply abandoned he
image py amid.
Figu e[3.2.1.3] (c), SSD ied o use CNN py amid-shaped hie a chical ea u es
ea lie . Ideally, he SSD-s yle py amid euses he mul i-scale ea u e maps om mul iple
laye s calcula ed by he o wa d p ocess, so his o m does no consume addi ional
esou ces. Howe e , in o de o a oid he use o low-le el ea u es, SSD abandoned he
shallow ea u e map, bu s a ed building py amids om con 4_3, and added some new
laye s. The e o e, SSD has gi en up on eusing highe - esolu ion ea u e maps, bu hese
ea u e maps a e e y impo an o de ec ing small a ge s. This is he di e ence
be ween SSD and FPN.
Figu e[3.2.1.3] (4) is he s uc u e o FPN. FPN is a py amid o m o na u al use o
CNN hie a chical ea u es, while gene a ing ea u e py amids wi h s ong seman ic
in o ma ion a all scales. The e o e, he s uc u e o FPN is designed wi h a op-down
s uc u e and a ho izon al connec ion o in eg a e a shallow laye wi h high esolu ion
and a deep laye wi h ich seman ic in o ma ion. In his way, i is possible o quickly build
a ea u e py amid wi h s ong seman ic in o ma ion on all scales om a single inpu
image a a single scale, wi hou incu ing signi ican cos s.
FPN is no a comple e a ge de ec ion ne wo k, bu a ea u e py amid ne wo k.
The FPN men ioned is ac ually a ea u e py amid ex ac ion ea u e. The e o e, he idea
o Fas e RCNN plus FPN essen ially changes he ea u e ex ac ion pa because he e
a e mo e ea u e laye s. So ROIpooling has also inc eased.

16
Figu e[3.2.1.4] as e _ cnn+ pn s uc u e
① The backbone ne wo k gene a es ou -scale ea u e maps, and hen
sequen ially passes h ough hei espec i e La e al_con o make hei channel numbe s
consis en .
②The deep ea u e maps a e down-sampled o adjacen laye scales s ep by
s ep, and hen he wo a e added and used.
③Ou pu he ea u e map a e usion
The low-le el ea u e seman ic in o ma ion is ela i ely small, bu he a ge
loca ion is accu a e; he high-le el ea u e seman ic in o ma ion is iche , bu he a ge
loca ion is ela i ely ough. Fpn is independen ly p edic ed in di e en ea u e maps.
17
Figu e[3.2.1.5] Independen p edic ion in di e en ea u e maps
Chap e 3.2.2 - One s age
1. SSD[6]
SSD uses he idea o meshing, and unlike Fas e RCNN, i in eg a es all ope a ions
in o a con olu ional ne wo k. In o de o de ec a ge s o di e en scales, SSD
pe o ms sliding window scanning on he ea u e images o di e en con olu ional
laye s; small a ge s a e de ec ed in he ea u e images ou pu by he p e ious
con olu ional laye , and la ge a ge s a e de ec ed in he ea u e images ou pu
by he subsequen con olu ional laye . The goal. I s main ea u es a e:
1.1. De ec ion based on mul i-scale ea u e images: P edic ion on mul i-scale
con olu ion ea u e maps o de ec a ge s o di e en sizes, which imp o es he
de ec ion accu acy o small a ge objec s o a ce ain ex en .
1.2. D awing on he idea o Ancho boxes in Fas e R-CNN, sampling candida e
egions on ea u e maps o di e en scales, which imp o es he ecall a e o
de ec ion and he de ec ion e ec o small a ge s o a ce ain ex en . The
ollowing igu e shows he p inciple o SSD:
18
Figu e[3.2.2.1] Schema ic o SSD
2. VGG16[8]
VGG was p oposed in 2014 by he Visual Geome y G oup, Depa men o
Science and Enginee ing, Ox o d Uni e si y. The main wo k is o p o e ha
inc easing he dep h o he ne wo k can a ec he inal pe o mance o he
ne wo k o a ce ain ex en . VGG has wo s uc u es, namely VGG16 and VGG19.
Excep o he di e ence in ne wo k dep h, he e is no di e ence in essence
be ween he wo. Compa ed wi h AlexNe in 2012, a high ad ance o VGG is o
use con inuous 3x3 small con olu ion ke nels o eplace he la ge ones in AlexNe
(AlexNe uses 11x11, 7x7 and 5x5 con olu ion ke nels). The supe posi ion o wo 3x3
con olu ion ke nels wi h a s ep leng h o 1, i s ecep i e ield is equi alen o a 5x5
con olu ion ke nel. Howe e , he use o s acked small con olu ion ke nels is due
o he la ge con olu ion ke nel, because he inc ease in he numbe o laye s
19
inc eases he nonlinea i y o he ne wo k, which allows he ne wo k o lea n mo e
complex models, and he small con olu ion ke nel has ewe pa ame e s.
Figu e[3.2.2.2] VGG16 model
VGG16 con ains:
13 con olu ional laye s (Con olu ional Laye ), espec i ely ep esen ed by con 3-
XXX
3 ully connec ed laye s (Fully connec ed Laye ), espec i ely ep esen ed by FC-
XXXX
5 pooling laye s (Pool laye ), espec i ely ep esen ed by maxpool
The ou s anding ea u e o VGG16 is simplici y, which is e lec ed in:
3. The con olu ional laye s all use he same con olu ion ke nel pa ame e s
The con olu ional laye s a e all exp essed as con 3-XXX, whe e con 3 indica es
ha he ke nel size o he con olu ional laye used by he con olu ional laye is 3,
ha is, he wid h and heigh a e 3, and 3*3 is e y Small con olu ion ke nel size,
combined wi h o he pa ame e s (s ide=1, padding=same), so ha each
con olu ional laye ( enso ) can main ain he same wid h and wid h as he
26
a)Combine he o iginal Da kne 53 wi h CSPNe . YOLO 3 is composed o a se ies
o esidual s uc u es. A e he combina ion, he main job o CSPne is o spli he s ack o
he o iginal esidual block, and spli i in o wo pa s: he main pa con inues o s ack he
o iginal esidual block, and he b anch pa is equi alen o a esidual edge. A e a small
amoun o p ocessing, i is di ec ly connec ed o he end.
b)Use he MIsh ac i a ion unc ion o eplace he o iginal Leaky ReLU. In YOLO 3,
each con olu ional laye includes a ba ch no maliza ion laye and a Leaky ReLU. In he
backbone ne wo k CSPDa kne 53 o YOLO 4, Mish is used ins ead o he o iginal Leaky
ReLU.
(2)SPP[11]
The o iginal design pu pose o SPP is o make he con olu ional neu al ne wo k no
es ic ed by he ixed inpu size. In YOLO 4, he au ho in oduced SPP because i
signi ican ly inc eases he ecep i e ield, isola es he mos impo an con ex ea u es,
and ha dly educes he unning speed o YOLO 4. As shown in he igu e below, i is he
classic spa ial py amid pooling laye in SPP.
Figu e[3.2.4.3] spp space py amid pooling laye
In YOLO 4, he speci ic me hod is o use he maximum pooling o ou di e en scales o
p ocess he ea u e map ou pu by he uppe laye . The maximum pooling co e size is
13x13, 9x9, 5x5, 1x1, and 1x1 is equi alen o no p ocessing.

27
(3)PANe [12]
PANe as a whole can be seen as a numbe o imp o emen s on Mask R-CNN, making
ull use o ea u e usion, such as in oducing Bo om-up pa h augmen a ion s uc u e,
making ull use o ne wo k shallow ea u es o segmen a ion; in oducing Adap i e
ea u e pooling o make he ex ac ed ROI Fea u es a e mo e abundan ; Fully-
conneFc ed usion is in oduced, and a mo e accu a e segmen a ion esul is ob ained
by using he ou pu o a on -backg ound bina y classi ica ion b anch.
Figu e[3.2.4.4] PANe schema ic
(a)FPN backbone,(b)Bo om-up pa h augmen a ion,(c) Adap i e ea u e pooling,(d)Box
b anch,(e)Fully-connec ed usion
In YOLO 4, he au ho uses PANe ins ead o FPN in YOLO 3 as he me hod o pa ame e
agg ega ion, and pe o ms pa ame e agg ega ion om di e en backbone laye s o
di e en de ec o le els. And modi ied he o iginal PANe me hod, using enso
connec ion (conca ) ins ead o he o iginal sho cu connec ion (sho cu connec ion).
(4) YOLO 3 Head
In YOLO 4, he Head ha inhe i s YOLO 3 pe o ms mul i-scale p edic ion, which
imp o es he de ec ion pe o mance o a ge s o di e en sizes.YOLO 4 lea ns he
28
YOLO 3 me hod, uses h ee di e en le els o ea u e maps o usion, and inhe i s he
Head o YOLO 3.
(5)T icks
To ob ain be e accu acy wi hou inc easing he in e ence cos , bu only change he
aining s a egy o only inc ease he aining cos me hod, he au ho calls i "Bag o
eebies"; i only inc eases a small amoun o in e ence cos bu can signi ican ly imp o e
he accu acy o a ge de ec ion Plug-in modules and pos -p ocessing me hods, called
"Bag o specials"
- Bag o eebies
Random zoom Flip, o a e Image dis u bance, noise, occlusion Change b igh ness,
con as , sa u a ion, andom e ase(Cu ou MixUp Cu Mix)
Common egula iza ion me hods a e: D opOu D opConnec D opBlock
The me hods o balance posi i e and nega i e samples a e: Focal loss OHEM (online
ha d- o-sepa a e sample mining)
In addi ion, he e a e imp o emen s in e u n loss: GIOU DIOU CIOU
- Bag o specials
Inc ease he ecep i e ield skills: SPP ASPP RFB
A en ion mechanism: Squeeze-and-Exci a ion (SE) Spa ial A en ion Module (SAM)
Fea u e usion in eg a ion: FPN SFAM ASFF BiFPN ( om he amous E icien De )
Be e ac i a ion unc ion: ReLU LReLU PReLU ReLU6 SELU Swish ha d-Swish
Pos -p ocessing non-maximum supp ession algo i hm: so -NMS DIoU NMS
(6)Ways o imp o e
In addi ion o he a ious T icks men ioned abo e, in o de o make he a ge de ec o
easie o ain on a single GPU, he au ho also p oposes 5 imp o ed me hods:
29
- Mosaic
This is a new da a enhancemen me hod p oposed by he au ho , which d aws on he
idea o Cu Mix da a enhancemen me hod. Cu Mix da a enhancemen me hod uses
wo pic u es o s i ching, bu Mosaic uses ou pic u es o s i ching.
- SAT
SAT is a sel -ad e sa ial aining da a enhancemen me hod, which is a new ad e sa ial
aining me hod. In he i s s age, he neu al ne wo k changes he o iginal image wi hou
changing he ne wo k weigh s. In his way, he neu al ne wo k conduc s an ad e sa ial
a ack on i sel , changing he o iginal image o c ea e a decep ion ha he e is no
desi ed objec on he image. In he second s age, he no mal me hod is used o ain he
neu al ne wo k o de ec he a ge .
- CmBN
The ull name o CmBN is C oss mini-Ba ch No maliza ion, which is de ined as C oss Mini-
Ba ch No maliza ion (CmBN). CmBN is an imp o ed e sion o CBN, which is used o
collec s a is ical da a in mul iple mini-ba ches in a ba ch.
- SAM
The au ho modi ied he o iginal SAM (Spa ial A en ion Module) me hod, changing SAM
om spa ial a en ion o poin a en ion. As shown in he igu e below, o he
con en ional SAM, he maximum pooling laye and he a e age pooling laye ac on
he inpu ea u e maps espec i ely o ob ain wo se s o ea u e maps wi h he same
shape, and hen inpu he esul s in o a con olu ional laye , ollowed by Sigmoid unc ion
o c ea e spa ial a en ion.
- PAN
The au ho modi ied he o iginal PAN (Pa h Agg ega ion Ne wo k) me hod, using enso
connec ion (conca ) ins ead o he o iginal sho cu connec ion (sho cu connec ion)
30
Chap e 4 - Da a Se
Chap e 4.1 - Tshingua-Daimle [13]
Tsinghua-Daimle Cyclis De ec ion con ains 9741 images wi h anno a ions only o
"cyclis ". Only cyclis s which a e ully isible (occlusion<10%) and highe han 60 pixels
ha e been labeled he e.
Chap e 4.2 - INRIA Pe son Da ase [14]
This da ase has c opped he image o pedes ians so ha he pe son can be
mo e p ominen , and he pos u e is no pa icula ly s ange. Only a ew pic u es a e
ob ained om Google.
Chap e 4.3 - C ea ed da a se
I did no selec pedes ians om Tsinghua's da a se , because he size o
pedes ian images in Tsinghua's da a se is 64x64, which is oo small o expe imen s, so I
did no choose. Fo he KITTI da a se , when I go hei da a se , I ound ha i was all ca s,
so I judged ha i did no mee my opic.
Finally.
Cyclis da a se adop s Tsinghua-Daimle Cyclis De ec ion pa o he da a se ,1507
image esolu ion 2048*1024.
The Pe son da a se uses he 614 aining se and 288 es se o he INRIAPe son pedes ian
da a se .
A o al o 2088 shee s o all da a se s a e di ided in o aining se (1671) and es se (417)
a a a io o 8:2.
31
All he pic u es used need o use he .jpg o ma . The c ea ed da a se uses codes o
dis inguish he aining se and he alida ion se , and hen uses he code o con e he
VOC o ma .
They a e
1.Anno a ions (.xml, each pic u e co esponds o a ile, which s o es he name o he
pic u e, he long sec ion, he uppe le and lowe igh coo dina es o he a ge ame,
he ca ego y name o he a ge ame)
2.ImagesSe (. x , sa e The name o he image ha needs o be ained)
3.JPEGImage ( he o iginal image)
Chap e 5 - Ha dwa e
The ha dwa e I used o comple e his p ojec . Knowing his is necessa y, because
di e en ha dwa e will a ec he aining p ocess, especially he aining ime, o
e iciency.
Chap e 5.1 - In el(R) Co e(TM) i7-9700F CPU
l Co es: 8
l Th eads: 8
l P ocesso Base F equency: 3.00GHz
l Max Tu bo F equency: 4.7GHz
l TDP: 65W

32
Chap e 5.2 - GEFORCE RTX 2080 Ti
l NVIDIA CUDA ® Co es: 4352
l Boos Clock (MHz): 1545
l Base Clock (MHz): 1350
l S anda d Memo y Con ig: 11 GB GDDR6
l G aphics Ca d Powe (W): 250W
Chap e 5.3 - Je son Nano 2GB De elope Ki [28]
l CPU Quad-co e ARM® A57 @ 1.43 GHz
l GPU 128-co e NVIDIA Maxwell™
l MEMORY 2 GB 64-bi LPDDR4 25.6 GB/s
Since i s elease in 2019, NVIDIA Je son Nano has c ea ed a enzy in he ield o AIOT
edge compu ing applica ions wo ldwide, He was also named "Bes AI P ocesso " in he
2020 Bes Visual P oduc s awa d lis by Edge AI and Vision Alliance. This is mainly because
he NVIDIA Je son Nano chip ully mee s he six challenges o AIOT chip: high
pe o mance, small size, low powe consump ion, ull in e ace, ecological in eg i y, and
excellen cos .
The eason o using his edge de ice was o es a ious models on a low-powe machine.
Fo compa ison, es he FPS on a low-pe o mance machine.
33
Chap e 6 - Ta ge de ec ion e alua ion
The classi ica ion ask will be judged by he p ecision/ ecall cu e. The p incipal
quan i a i e measu e used will be he a e age p ecision (AP).[15] The e a e some
impo an hings o know abou calcula ing AP.[16]
Chap e 6.1 - In e sec ion O e Union (IOU)
Figu e[6.1.1] IOU o mula
To pu i simply, we equi e ha he IOU o he wo boxes is di ided by he
o e lapping pa o he wo boxes by he union o he wo boxes. I is clea om he
second o mula abo e. Pay a en ion o he wo boxes when calcula ing. The a ea o
he union is equal o he sum o he a eas o he wo minus he a ea o he in e sec ion o
he wo.
Chap e 6.2 - T ue Posi i e, False Posi i e, False Nega i e and T ue
Nega i e
T ue Posi i e (TP): A co ec posi ioning esul is ha he IOU be ween you
p edic ed box and ou g ound u h can be g ea e han ou speci ied h eshold, we
gene ally ake his h eshold o be 0.5
34
False Posi i e (FP): I is a w ong esul , ha is, he IOU o he box and g ound u h
you p edic ed is less han he h eshold
False Nega i e (FN): We o iginally had an objec he e, so he e should be a box in
ou place, bu you did no p edic i , hen his g ound u h is a FN o you model
T ue Nega i e (TN): This is ha ou g ound u h has ound a p edic ion box whose
IOU is g ea e han he h eshold, hen he g ound is conside ed o be success ully
de ec ed
Chap e 6.3 - P ecision
P ecision is he abili y o a model o iden i y only he ele an objec s. I is he
pe cen age o co ec posi i e p edic ions and is gi en by:
Chap e 6.4 - Recall
Recall is he abili y o a model o ind all he ele an cases (all g ound u h
bounding boxes). I is he pe cen age o ue posi i e de ec ed among all ele an
g ound u hs and is gi en by:
35
Chap e 6.5 - AP calcula ion
(1) Be o e VOC2010, you only need o selec he maximum P ecision when
Recall >= 0, 0.1, 0.2, ..., 1 o al 11 poin s, and hen AP is he a e age o hese 11 P ecision.
Figu e[6.5.1] AP calcula ion
Calcula ing he o al a ea, we ha e he AP.
(2) In VOC2010 and la e , o each di e en Recall alue (including 0 and 1), selec
he maximum P ecision when i is g ea e han o equal o hese Recall alues, and hen
calcula e he a ea unde he PR cu e as he AP alue. Finally he a e age AP alue o
all classes is mAP.
Chap e 7 - F amewo ks
Chap e 7.1 - Cuda
CUDA (an ac onym o Compu e Uni ied De ice A chi ec u e) is a pa allel compu ing
pla o m and applica ion p og amming in e ace (API) model c ea ed by N idia.[17]I
allows so wa e de elope s and so wa e enginee s o use a CUDA-enabled g aphics
42
Figu e[8.3.2] loss line cha o YOLO 4
The same is a single-laye a ge de ec ion, yolo 4's pe o mance is signi ican ly
be e han be o e.
Chap e 8.4 - Resul o FASTER RCNN+FPN

43
Figu e[8.4.1] mAP line cha o FASTER RCNN+FPN
Figu e[8.4.2] loss line cha o FASTER RCNN+FPN
I am e y sa is ied wi h he pe o mance o as e cnn+ pn, because he same
da a se has a e y big di e ence h ough he algo i hm. Below I will compa e all he
abo e esul s oge he .
Chap e 9 - Compa ison o expe imen al esul s
Chap e 9.1 - Compa ison o esul s o high-pe o mance machines
Model
Inpu Size
FPS
mAP
Cyclis
Pe son
VGG16
512*512
47.3
0.81
0.85
0.77
MobileNe 3
320*320
43.1
0.67
0.54
0.80
YOLO 4
608*608
19.6
0.78
0.75
0.81
FASTER
RCNN+FPN
1333*800
12
0.9
0.899
0.902
Table[9.1] Compa ison o algo i hm esul s-high
44
The able shows us he a ious da a o each algo i hm in he high-pe o mance
compu e .
F om he expe imen al esul s, we can see ha FASTER RCNN+FPN pe o ms bes in
e ms o de ec ion accu acy, bu he sho comings a e also e y ob ious. The ps only
eached 12. This is due o i s wo-s age ne wo k. I has a egion p oposal ne wo k RPN, his
pa uses CNN, and he egion p oposal ne wo k and classi ica ion ne wo k sha e
con olu ional ea u es. And he e is FPN, independen p edic ion o di e en ea u e
laye s. These can e ec i ely a ge de ec ion, bu he speed is slowe . We can easily ind
his om he able.
The abo e SSD-MobileNe inpu esolu ion is ela i ely small, and he ecogni ion
e ec o he pe son class wi h he o iginal image size o abou 800 is be e . The inpu
esolu ion o SSD-VGG16 is ela i ely la ge, and i can be e ecognize he cyclis class
wi h he o iginal image size o 2048*1024.
The eason why mobileNe 's mAP is oo low is because o o e i ing in he pe son
pa . I is because he da a se o he pe son pa is oo small. The way o sol e his
p oblem and ge he pe o mance i should be is o inc ease he amoun o da a, as I
explained in he esul s sec ion.
YOLO 4 is al eady e y good as a non-ligh weigh pe o mance. I shows a e y
a e age alue, and he e a e some aspec s ha mus be explained. Compa ed wi h i s
p e ious e sion, i has added many imp o emen s, such as adding ea u e py amids
(SPP, PAN) o he las con olu ional laye , and modi ying he ac i a ion unc ion o Mish,
lea ning a e annealing algo i hm. F om he s uc u al poin o iew, YOLO 4 also has an
FPN s uc u e. Compa ed wi h FASTER RCNN+FPN, YOLO 4 has a highe ps pe o mance.
Al hough mAP did no each he wo s age, I go a be e esul in he one s age.
45
Chap e 9.2 - Compa ison o esul s o low-pe o mance machines
Model
FPS
VGG16
0.98
MobileNe 3
7.61
YOLO 4
--
FASTER RCNN+FPN
--
Table[9.2] Compa ison o algo i hm esul s-low
On low-pe o mance compu e s, I ocus on he pe o mance o ps. Because he
machine I used is he je son nano 2gb e sion o NVIDIA. I means ha in he case o
suppo ing a se ies o en i onmen s, he heo e ical ideo memo y o he machine is only
2gb. In ac ual use, he ideo memo y is only abou 1.3gb. This has ac ually eached a
limi alue. F om he able, we can see ha in he case o e y low ideo memo y, only
ligh weigh algo i hms can ac ually un no mally and pe o m a ge ecogni ion. The
no mal e sion o YOLO 4 canno un on a machine wi h only 2gb o ideo memo y. The
wo-s age FASTER RCNN+FPN is e en mo e impossible.
The e o e, o a low-pe o mance machine, i is e y impo an o choose a co ec
algo i hm. F om he able, as a simple con olu ional neu al model, VGG16 can ope a e
co ec ly and ecognize a ge s no mally. Bu I hink i s ps is no good enough. As a
ligh weigh neu al ne wo k, MobileNe can un no mally and has a ps o abou 7 I hink i
is unde s andable. The e may be doub s ha his ps is no e y high, why is i s ill a
ligh weigh neu al ne wo k? This is because we a e using he La ge e sion, which has 3
46
bo leneck laye s and a poin -wise con olu ional laye mo e han he Low e sion. I s
Dep hwise Sepa able Con olu ion is e y sui able o low-pe o mance e sions.
Chap e 10 - Conclusions and u u e wo k
Chap e 10.1 - Conclusions
Fo hese ou algo i hms and he selec ed da a se , om he esul s, i he machine
pe o mance is no conside ed, hen FASTER RCNN+FPN is undoub edly he bes auxilia y
sys em. Because i has highe accu acy. I is undeniable ha wo s age is be e han one
s age in e ms o accu acy. Bu he sho comings a e also e y ob ious. The aining ime
o he wo s age is much longe han ha o he one s age. And in ac ual use, he ps o
he wo s age applica ion on high-pe o mance machines is no high enough, and i
canno un smoo hly on low-pe o mance machines.
The speed o yolo 4 is much as e han as e - cnn on high-pe o mance
machines. And because he ea u e ex ac ion laye o yolo 4 uses a ea u e py amid
s uc u e, i can also ge e y good accu acy.
SSD-VGG16 uses ea u e maps o di e en scales o de ec ion. This may also be
he eason why in his da a se , o small a ge s such as people and bicycles, and he e
may be obs uc ions on he oad, i has a highe accu acy a e han he YOLO algo i hm.
Fo his ime, MobileNe 3, which pe o med bes in low-pe o mance machines. I
hink he ligh weigh neu al ne wo k can be used as a solu ion in sol ing business p oblems,
o in ehicle sa e y sys ems. Because he speed o he ca is as and he neu al ne wo k
needs o be as , hen a ligh weigh neu al ne wo k is a e y good choice. Video memo y,
memo y, and cpu will be ela i ely small. O cou se, his is jus a s uc u al poin o iew o
hese ou algo i hms.
47
I hink he choice o he aining da a se is e y impo an . F om he inal esul , he
choice o he pic u e size is ac ually e y impo an , because he e a e ac ually
es ic ions on he pic u es on he In e ne o he pic u es aken by you sel . Today's
came a equipmen is ge ing be e and be e , o example, he pixels o mobile phones
a e ge ing be e and be e . The equi emen s o he aining se a e also e y high.
Maybe a ce ain a ge de ec ion algo i hm is excellen , bu due o he inapp op ia e
choice o he da a se , he inal esul is bad. So in he end, because o he di e en
models o di e en models and he di e en came a op ions in he ac ual use, he inal
esul s will appea o be di e en .
I inally used all he algo i hms o make ac ual de ec ions o assis ed d i ing. Using
he ideo o lea ning o d i e in a d i ing school, a no ice may no no ice pedes ians
and bicycles on he oad. 4 algo i hms a e used o he sc eensho s o he ideo in he
appendix. I am sa is ied wi h he aining esul s and can achie e a ge de ec ion o
mos o he scenes.
Chap e 10.2 - Fu u e Wo k
Fo u u e wo k, om he a ge de ec ion amewo k. F om he Fas e R-CNN
ela ed se ies ocusing on accu acy o he YOLO se ies ocusing on speed, he u u e
a ge de ec ion esea ch di ec ion will ocus mo e on he balance o accu acy and
speed. The e o e, many ne wo k s uc u es a e gene a ed on he SSD amewo k. I hink
ha i a ge de ec ion is o be used as an in- ehicle sa e y sys em, a ious ligh weigh
neu al ne wo ks a e he mos p e e ed in ecen yea s. The e a e epo s ha YOLO 4-
iny pe o ms e y well, we can also y i in low-pe o mance machine. The new YOLO 5
has appea ed, we can y o use i in high-pe o mance machine.

48
I ha e seen some esea ch pape s ha show ha he o a ion in a iance o
con olu ional neu al ne wo ks is ob ained h ough da a enhancemen and la ge sample
lea ning, and i does no ha e his ea u e i sel . Maybe con inue o s udy his will change
he basic amewo k o he a ge de ec ion algo i hm.
The use o 3D lida will be mo e e ined han he image de ec ion. Be e senso s
will make a ge de ec ion mo e e ined, and now mo e au onomous d i ing applica ions
on he ma ke a e 3D lida .
O cou se, some con en ional ideas a e o imp o e he loss unc ion, ac i a ion
unc ion, and ea u e usion, which can also be conside ed.
49
BIBLIOGRAPHY
[1] Good ellow, I., Bengio, Y., Cou ille, A.．Deep lea ning (Vol. 1)．Camb idge：MIT p ess，2016
[2] Gu, J., Wang, Z., Kuen, J., Ma, L., Shah oudy, A., Shuai, B., Liu, T., Wang, X., Wang, L., Wang, G. and Cai, J.,
2015. Recen ad ances in con olu ional neu al ne wo ks. a Xi p ep in a Xi :1512.07108.
[3] Zhang, W., 1988. Shi -in a ian pa e n ecogni ion neu al ne wo k and i s op ical a chi ec u e. In P oceedings o
annual con e ence o he Japan Socie y o Applied Physics.
[4] Shaoqing Ren, Kaiming He, Ross Gi shick, and Jian Sun.Fas e R-CNN: Towa ds Real-Time Objec De ec ion
wi h Region P oposal Ne wo ks
[5] Tsung-Yi Lin1,2, Pio Dolla ´ 1 , Ross Gi shick1 , Kaiming He1 , Bha a h Ha iha an1 , and Se ge Belongie.Fea u e
Py amid Ne wo ks o Objec De ec ion
[6] Wei Liu1 , D agomi Anguelo 2 , Dumi u E han3 , Ch is ian Szegedy3 , Sco Reed4 , Cheng-Yang Fu1 ,
Alexande C. Be g.SSD: Single Sho Mul iBox De ec o
[7] Alexey Bochko skiy,Chien-Yao Wang,Hong-Yuan Ma k Liao.YOLO 4: Op imal Speed and Accu acy o Objec
De ec ion
[8] Simonyan,Zisse man.Ve y Deep Con olu ional Ne wo ks o La ge Scale Image Recogni ion
[9] And ew Howa d, Ma k Sandle , G ace Chu, Liang-Chieh Chen, Bo Chen,Mingxing Tan, Weijun Wang, Yukun Zhu,
Ruoming Pang, Vijay Vasude an, Quoc V. Le, Ha wig Adam,Sea ching o MobileNe V3
[10] Chien-Yao Wang,Hong-Yuan Ma k Liao,I-Hau Yeh,Yueh-Hua Wu,Ping-Yang Chen,Jun-Wei Hsieh,CSPNET: A
NEW BACKBONE THAT CAN ENHANCE LEARNING CAPABILITY OF CNN
[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun,Spa ial Py amid Pooling in Deep Con olu ional
Ne wo ks o Visual Recogni ion
[12] Shu Liu, Lu Qi, Hai ang Qin, Jianping Shi, Jiaya Jia,Pa h Agg ega ion Ne wo k o Ins ance Segmen a ion
[13] Y. Yang H. Xiong M. B aun S. Pan K. Li X. Li, F. Floh and D. M. Ga ila. A new benchma k o ision-based
cyclis de ec ion, June 2016.
[14] Na nee DALAL,Finding People in Images and Videos
[15] Ma k E e ingham, John Winn, The PASCAL Visual Objec Classes Challenge 2012 (VOC2012) De elopmen Ki
[16] h ps://gi hub.com/ a aelpadilla/Objec -De ec ion-Me ics
[17] h ps://de elope .n idia.com/cuda-zone
[18] Abi-Chahla, Fedy (June 18, 2008). "N idia's CUDA: The End o he CPU?". Tom's Ha dwa e. Re ie ed May 17,
2015.
[19] Yegulalp, Se da (19 Janua y 2017). "Facebook b ings GPU-powe ed machine lea ning o Py hon". In oWo ld.
Re ie ed 11 Decembe 2017
[20] Lo ica, Ben (3 Augus 2017). "Why AI and machine lea ning esea che s a e beginning o emb ace PyTo ch".
O'Reilly Media. Re ie ed 11 Decembe 2017
50
[21] Ke ka , Nikhil (2017). "In oduc ion o PyTo ch". Deep Lea ning wi h Py hon. Ap ess, Be keley, CA. pp. 195–208.
doi:10.1007/978-1-4842-2766-4_12. ISBN 9781484227657.
[22] Pa el, Mo (2017-12-07). "When wo ends use: PyTo ch and ecommende sys ems". O'Reilly Media. Re ie ed
2017-12-18.
[23] Mannes, John. "Facebook and Mic oso collabo a e o simpli y con e sions om PyTo ch o Ca e2". TechC unch.
Re ie ed 2017-12-18. FAIR is accus omed o wo king wi h PyTo ch – a deep lea ning amewo k op imized o
achie ing s a e o he a esul s in esea ch, ega dless o esou ce cons ain s. Un o una ely in he eal wo ld,
mos o us a e limi ed by he compu a ional capabili ies o ou sma phones and compu e s.”
[24] A akelyan, Sophia (2017-11-29). "Tech gian s a e using open sou ce amewo ks o domina e he AI communi y".
Ven u eBea . Re ie ed 2017-12-18.
[25] "The C++ F on end". PyTo ch Mas e Documen a ion. Re ie ed 2019-07-29.
[26] h ps://gi hub.com/open-mmlab/mmc /blob/mas e /README.md
[27] SAE. Taxonomy and de ini ions o e ms ela ed o d i ing au oma ion sys ems o on- oad mo o ehicles.
h ps://www.sae.o g/misc/pd s/au oma ed_d i ing.pd ,2018.
[28] h ps://de elope .n idia.com/embedded/je son-nano-2gb-de elope -ki
51
Appendix
Table[1] SSD-MobileNe mAP
Epoch
mAP
Cyclis
pe son
1
0.612
0.425
0.798
2
0.649
0.491
0.807
3
0.658
0.504
0.811
4
0.669
0.540
0.799
5
0.657
0.499
0.814
6
0.652
0.498
0.805
7
0.663
0.519
0.808
8
0.648
0.494
0.802
9
0.668
0.534
0.803
10
0.665
0.533
0.797
11
0.670
0.541
0.798
12
0.667
0.537
0.797