[DL輪読会]Tracking Objects as Points

DEEP LEARNING JP [DL Papers] Tracking Objects as Points Shizuma Kubo, ACES.Inc http://deeplearning.jp/ 1

• http://deeplearning.jp/

• mhn Ø 2I CA Ø tj 0: A A Ø l GL G Ø G KL K 1 A AHH I :L / : ) Xa . GIE Kv b] 4 I A ( Ø io s 1GA K 2 (L KA Ø Ø K c ( ( / / ))4 rXfgXVOep uk ry [TO P AK,L:rXZXU ud

• https://github.com/xingyizhou/CenterTrack

• ( ( Ø Ø e Ø Ø • n 136 ) ( ( (s ) D il A: P M t woSa E M A7 7E - - J J 7 E2 u . r D0 7E / ( :AC 7 DC P Mjgl L J CE e 36 h fc djgl h fc il djgl O p MN TJ136 M b

• https://www.slideshare.net/DeepLearningJP2016/dlobjects-as-points

.5 .5 32 1 4

Ø Ø 5 Ø . 4 4 . C 32

Multi Object Tracking (MOT) • • Ro l Ø k c jTmd W i l Ø Ø Sr awT] Ø • jTmd p p W DS I O S[ M S g Tt jTne Sr DS st S O r g h +6: 6 2 6: 1 6 A6 16 6 61 7

MOT Challenge • O i N P ( 177 x MP O M(0- M • lO • N 2 Ø 20 N 1 2 • • 1 2 2 2 20 Or Ø 199 e V 2 O 20 20 N g )1 MPF s T Cn ( d i Os r r O O V : pR ah 2 M M O O tVO N V M VFN i

• ) a o % • e ) e e • (0 1 • (0 1 • ( gT D : P gs v S • • i ( ISNI T G T D W o G F : g s g D o g 81 120 81 F D oT P tF D l %

MOT • - - ng 9 : Ø a Mt Ø M b • M r M i c 9 : : P 9 T M9 PO - - - - : ng M i Ø c : u e] Ø ( Pv e] ) ng [ D : 9 xM ] do l b M T9

MOT : Tracking-by-detection • 1 6 Ø [r ) c Ø Tr • jdi 0 ]B u eBa a gj ʼ 1 Ø c Ø • gj a [r (- [k ]B I [r n 01 1 i ow t b W UW yx C P lx c ]B gj

MOT : Joint detection and tracking • VekFhlF Ø 9 Ø 8 Ø + • • 98 B8 7 ]Z V 9 7 1 V 7 C D B78 ekFh sTS nt 8 7 IgFb ʼ Vdic V ʻ I h C V oI uT wr ] P I [amf Tt n

)( ) 0 2 • • nr 6 27 6 1 0 Emr ) D v c ea ) D ( • 61 70 ) M c A 0 706 6 2t D C • ) D o P TD k O V i

Limitation • W TO S 3 f • o7 3 TO D 3 S f I6 • L 3 k 11 e M

Ø Ø . . . 435 1

CenterTrack • Our tracker, CenterTrack, applies a detection model to a pair of images and detections from the prior frame. Given this minimal input, CenterTrack localizes objects and predicts their associations with the previous frame. That’s it.

CenterTrack • 1 (1( 1 a f • a Ø Ø e Tc d i g a J a 6 kT )) (1 6 e

CenterNet (Objects as Points) • C + )22 • C a b a 7=c 1 + 24 C ( L height • e 1+) 1 2 width : https://medium.com/machine-learning-bites/deeplearning-series-convolutional-neural-networks-a9c2f2ee1524

• https://medium.com/machine-learning-bites/deeplearning-series-convolutional-neural-networks-a9c2f2ee1524

CeterTrack • N e ( • ( 2 1) 2 :C ( • ( • N N e 1 2 1) 2 2 1) 2 8 3 + C 2 1) 2 4 :C ( :Cd 3 + 2 1) 2 4 b

15 Ø Ø 9 25 . 4

+ ( • e CN : + ) 1 2 Ø : Ø : • 01 :

(offset prediction) • ri c Ø no ) Ø , Ø N2 Ø Ø +() )1 +) 1() f )d t 1 ) ) ) e )1 s ) e c : c c c C ) 1 ) ) ) )

Ablation study • / • l Ø • 2 o h gp i ar N mD C : f e h Ø n i SW e D I : f

offset prediction ( • l ) t c R R 2T • f p e Ø i p O S R 2 c a R 2T : st 3 Ø Ø 2T w : o

315 315 Ø Ø .2 4

• F 5 m • m 5 Fi N t v :F Ø • n o N F F P e A 2 F Ø ( ) Ø ( ) 5 ) s u 5 l A ) ( ) s 5 Fg v F 5 a F ) (

• a e Ø Ø [ p [ ] 2 2() i c • T Ø l s Ø k a e [Tf c g : r Tyu vn 2 2 ) 6 Ø tn t [ ] T C 2() [ ]

Ablation study • a Ø i ) ( Ø 3 du t 3 • ) ( m Ø • r : . 2 : e r o Ø n Ø ) ( ri : r 3 A 7e a ) (( g : e

. . 2 43 15

• 2 Ø 2 Ø • C Ø Ø • ( ) : I 2 oe 9 D 2 D C s f