๐Ÿ›‘

CenterNet: Objects as Points

Tags
Object Detection
Created
2021/01/29 07:01
Publication
Rate
3
Source
https://arxiv.org/abs/1904.07850
Summary
๊ต‰์žฅํžˆ ๋น ๋ฅด๋ฉด์„œ๋„ ์ค€์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” one-shot detector๋ฅผ ์ œ์•ˆํ•œ๋‹ค. (์ƒ์„ธ ํŽ˜์ด์ง€ ์ฐธ๊ณ )

Reference

โ€ข
โ€ข
์ฐธ๊ณ  ๋ธ”๋กœ๊ทธ: https://nuggy875.tistory.com/34

Introduction

2-stage object detection์˜ ๊ฒฝ์šฐ์—๋Š” ๋ฌผ์ฒด์˜ ์œ„์น˜๋ฅผ ๋จผ์ € ์ถ”๋ก ํ•œ ๋‹ค์Œ, ์ถ”๋ก ๋œ ์œ„์น˜๋“ค์— ๋Œ€ํ•œ ๊ฐœ๋ณ„์ ์ธ classification์„ ์ˆ˜ํ–‰ํ•œ๋‹ค. ์ด ๊ฒฝ์šฐ ํ•™์Šต์€ ๋ฌผ๋ก ์ด๊ฑฐ๋‹ˆ์™€ ์ถ”๋ก ์— ๊ฝค ์˜ค๋žœ ์‹œ๊ฐ„์ด ๊ฑธ๋ฆฌ๋Š” ์‹ฌ๊ฐํ•œ ๋ฌธ์ œ์ ์ด ๋ฐœ์ƒํ•œ๋‹ค. ํ•œํŽธ YOLO์™€ ๊ฐ™์€ ์ด์ „์˜ 1-sage object detection์˜ ๊ฒฝ์šฐ์—๋Š” ๋‹จ ํ•œ ๋ฒˆ์˜ forward propagation์œผ๋กœ ์ถ”๋ก ์ด ๊ฐ€๋Šฅํ•˜์ง€๋งŒ, ๋‹ค์ˆ˜์˜ anchor box๋ฅผ ์‚ฌ์šฉํ•จ์œผ๋กœ์จ 1) anchor box๋ฅผ ์ง์ ‘ ์„ ํƒํ•ด์ค˜์•ผ ํ•˜๊ณ , 2) ์—ฌ๋Ÿฌ anchor box์— ์˜ํ•ด ๋™์ผํ•œ ๋ฌผ์ฒด๊ฐ€ ์ธ์‹๋จ์œผ๋กœ์จ NMS(Non-Max Suppression)๋ผ๋Š” ํ›„์ฒ˜๋ฆฌ๋ฅผ ๋”ฐ๋กœ ๊ฐ€ํ•ด์ค˜์•ผ ํ•œ๋‹ค๋Š” ๋ฌธ์ œ๊ฐ€ ์žˆ๋‹ค.
์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋“ค์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด CenterNet์€ ๋ฌผ์ฒด๋ฅผ ๊ทธ ์ค‘์‹ฌ์ ์œผ๋กœ์จ ์ธ์‹ํ•˜๊ณ , box size๋‚˜ ๋‹ค๋ฅธ feature๋“ค(์˜ˆ๋ฅผ ๋“ค์–ด pose estimation์„ ์œ„ํ•œ joint๋“ค)์„ ์ด๋ฏธ์ง€๋กœ๋ถ€ํ„ฐ ์ง์ ‘์ ์œผ๋กœ regressํ•˜๋Š” ๋ฐฉ์‹์„ ํƒํ•œ๋‹ค. CenterNet์€ ์ค‘์‹ฌ์ ์„ ์ฐพ์•„๋‚ด๊ธฐ ์œ„ํ•ด ์ค‘์‹ฌ์ ์— ๋Œ€ํ•œ heatmap์„ ์ƒ์„ฑํ•˜๊ณ , ๊ทธ๋ ‡๊ฒŒ ์ƒ์„ฑ๋œ heatmap์˜ peak point(์˜ˆ์ปจ๋Œ€ ์ฃผ๋ณ€ 9 ๊ทธ๋ฆฌ๋“œ ์ค‘ ๊ฐ€์žฅ ๊ฐ’์ด ๋†’์€ ๊ทธ๋ฆฌ๋“œ)๋ฅผ ์ค‘์‹ฌ์ ์œผ๋กœ ์„ ํƒ(๊ทธ๋ฆฌ๊ณ  ๊ทธ ์ค‘์‹ฌ์ ์— ๋Œ€ํ•ด ๋‹ค๋ฅธ feature๋“ค์„ regress)ํ•˜๋Š”๋ฐ, ์ด๋กœ์จ ํ›„์ฒ˜๋ฆฌ๊ฐ€ ํ•„์š”ํ•˜์ง€ ์•Š์€ 1-stage detection์ด ๊ฐ€๋Šฅํ•˜๊ฒŒ ๋œ๋‹ค.
๊ธฐ๋ณธ์ ์œผ๋กœ๋Š” YOLO ๋“ฑ์˜ 1-stage detection๊ณผ ์ ‘๊ทผ๋ฒ•์ด ์œ ์‚ฌํ•œ ํŽธ์ธ๋ฐ, ๋…ผ๋ฌธ์—์„œ ์ œ์‹œํ•˜๋Š” ํ•ต์‹ฌ์ ์ธ ์ฐจ๋ณ„์ ์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค.
โ€ข
First, our CenterNet assigns the "anchor" based solely on location, not box overlap. We have no manual thresholds for foreground and background classification.
โ€ข
Second, we only have one positive "anchor" per object, and hence do not need Non-Maximum Suppression. We simply extract local peaks in the keypoint heatmap.
โ€ข
Third, CenterNet uses a larger output resolution (output stride of 4) compared to traditional object detectors. This eliminates the need for multiple anchors.
ํ•œํŽธ CornerNet(๋ฐ•์Šค์˜ ์ขŒ์ƒ, ์šฐํ•˜์ ์œผ๋กœ ๋ฌผ์ฒด๋ฅผ ์ธ์‹)์ด๋‚˜ ExtremeNet(์ƒํ•˜์ขŒ์šฐ ๊ทน๋‹จ์ ๊ณผ ์ค‘์‹ฌ์ ์œผ๋กœ ๋ฌผ์ฒด๋ฅผ ์ธ์‹) ๋“ฑ ์ ์œผ๋กœ์จ ๋ฌผ์ฒด๋ฅผ ์ธ์‹ํ•˜๋ ค๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์ด์ „์—๋„ ์žˆ์—ˆ์ง€๋งŒ, CenterNet์€ ์ด๋“ค๋ณด๋‹ค ๊ฐ„๋‹จํ•˜๋ฉด์„œ(?)๋„ ํšจ๊ณผ์ ์ธ detection์„ ์ž๋ž‘ํ•œ๋‹ค.

Preliminary

CenterNet์—์„œ๋Š” ์ด๋ฏธ์ง€ IโˆˆRWร—Hร—3I \in R^{W \times H \times 3}์˜ mapping Y^โˆˆ[0,1]WRร—HRร—C\hat{Y} \in [0, 1]^{\frac{W}{R} \times \frac{H}{R} \times C}๋ฅผ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์„ ๋ชฉ์ ์œผ๋กœ ํ•œ๋‹ค. ์ด ๋•Œ
โ€ข
W,HW, H๋Š” ๊ฐ๊ฐ ์ด๋ฏธ์ง€์˜ ๋„ˆ๋น„์™€ ๋†’์ด, RR์€ output stride์ด๋‹ค.
โ€ข
CC๋Š” keypoint type์ด๋‹ค. ์˜ˆ์ปจ๋Œ€ detection์—์„œ CC๋Š” category์˜ ์ˆ˜, pose estimation์—์„œ CC๋Š” ๊ด€์ ˆ์˜ ์ˆ˜(๋…ผ๋ฌธ์—์„œ๋Š” 17์„ ์‚ฌ์šฉ)๊ฐ€ ๋  ์ˆ˜ ์žˆ๋‹ค.
โ€ข
Y^=1\hat{Y} = 1์€ ํ•ด๋‹น ๊ทธ๋ฆฌ๋“œ๊ฐ€ keypoint์ž„์„ ์˜๋ฏธํ•œ๋‹ค.
ํ•™์Šต์„ ์œ„ํ•ด์„œ๋Š” ์ด๋ฏธ์ง€์— ์žˆ๋Š” ๊ฐ๊ฐ์˜ keypoint์˜ low-resolution ์ขŒํ‘œ๋ฅผ ๊ตฌํ•˜๊ณ , Gaussian Kernel์„ ์ด์šฉํ•ด ๊ฐ๊ฐ์˜ ๊ทธ๋ฆฌ๋“œ์˜ ๊ฐ’์„ ๊ตฌํ•œ๋‹ค. ์—ฌ๊ธฐ์„œ Gaussian Kernel์€ ํ•ด๋‹น ๊ทธ๋ฆฌ๋“œ๊ฐ€ (๊ฐ€์žฅ ์ธ์ ‘ํ•œ - ์ฆ‰ ๊ฐ™์€ class์˜ Kernel๊ฐ’์ด ๊ฒน์น  ๊ฒฝ์šฐ, ๊ฐ’์ด ํฐ ๊ฒƒ์„ ํƒํ•œ๋‹ค) ์ค‘์‹ฌ์ ๊ณผ ์–ผ๋งˆ๋‚˜ ๊ฐ€๊นŒ์šด์ง€๋ฅผ ์ธก์ •ํ•œ๋‹ค. ์ฆ‰ ์ค‘์‹ฌ์ ์— ๊ฐ€๊นŒ์šธ์ˆ˜๋ก ๊ทธ ๊ฐ’์ด ํฌ๊ณ , ๋ฉ€์ˆ˜๋ก ๊ทธ ๊ฐ’์ด ์ž‘๋‹ค. ์ด๋ ‡๊ฒŒ ๊ตฌํ•œ ๊ฐ’๋“ค์ด ํžˆํŠธ๋งต์ด ๋˜๋Š” ๊ฒƒ์ด๋‹ค.
Yxyc=exp(โˆ’(xโˆ’px~)2+(yโˆ’py~)22ฯƒp2)Y_{xyc} = exp( -\frac{ (x - \tilde{p_x})^2 + (y - \tilde{p_y})^2 }{ 2\sigma_p^2 } )
๊ทธ๋Ÿฌ๋ฉด ๋‹ค์Œ๊ณผ ๊ฐ™์€ Focal Loss๋ฅผ ์ด์šฉํ•ด ํžˆํŠธ๋งต์„ ํ•™์Šตํ•œ๋‹ค.
Lk=โˆ’1Nโˆ‘xyc{(1โˆ’Y^xyc)ฮฑlogโก(Y^xyc),ifย Yxyc=1(1โˆ’Yxyc)ฮฒ(Y^xyc)ฮฑlogโก(1โˆ’Y^xyc),otherwiseย L_k = -\frac{1}{N} \sum_{xyc} \begin{cases} (1 - \hat{Y}{xyc})^{\alpha} \log{( \hat{Y}{xyc} )}, & \text{if}\ Y_{xyc} = 1 \\ (1 - Y_{xyc})^{\beta} (\hat{Y}{xyc})^{\alpha} \log{( 1 - \hat{Y}{xyc} )}, & \text{otherwise} \end{cases}ย 
์œ„์˜ Focal Loss๋ฅผ ํ•ด์„ํ•ด๋ณด์ž๋ฉด,
โ€ข
์ฒซ์งธ ํ•ญ์€ ํ•ด๋‹น ๊ทธ๋ฆฌ๋“œ๊ฐ€ ์ค‘์‹ฌ์ ์ผ ๊ฒฝ์šฐ, Y^\hat{Y}๊ฐ€ 1์—์„œ ์–ผ๋งˆ๋‚˜ ๋ฉ€๋ฆฌ ๋–จ์–ด์ ธ ์žˆ๋Š”์ง€๋ฅผ ์ธก์ •ํ•˜๋ฉฐ,
โ€ข
๋‘˜์งธ ํ•ญ์€ ํ•ด๋‹น ๊ทธ๋ฆฌ๋“œ๊ฐ€ ์ค‘์‹ฌ์ ์ด ์•„๋‹ ๊ฒฝ์šฐ, Y^\hat{Y}๊ฐ€ 0์—์„œ ์–ผ๋งˆ๋‚˜ ๋ฉ€๋ฆฌ ๋–จ์–ด์ ธ ์žˆ๋Š”์ง€๋ฅผ ์ธก์ •ํ•˜์ง€๋งŒ, ๊ทธ ๊ฐ’์„ Kernel ๊ฐ’์— ๋Œ€ํ•ด ๋ณด์ •ํ•˜์—ฌ ๊ณ„์‚ฐํ•œ๋‹ค. ์ฆ‰ Kernel ๊ฐ’์ด ์ž‘์„์ˆ˜๋ก(์ค‘์‹ฌ์ ์—์„œ ๋ฉ€๋ฆฌ ๋–จ์–ด์ ธ ์žˆ์„์ˆ˜๋ก) 0์—์„œ ๋ฉ€๋ฆฌ ๋–จ์–ด์ ธ ์žˆ์Œ์œผ๋กœ์จ ๋ฐœ์ƒํ•˜๋Š” loss๊ฐ€ ์ปค์ง„๋‹ค.
์—ฌ๊ธฐ์„œ ฮฑ,ฮฒ\alpha, \beta๋Š” ๊ฐ๊ฐ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ, NN์€ ์ด๋ฏธ์ง€ II์˜ keypoint์˜ ์ˆ˜์ด๋‹ค. ๋…ผ๋ฌธ์—์„œ๋Š” ฮฑ=2,ฮฒ=4\alpha = 2, \beta = 4๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค.
ํ•œํŽธ ํžˆํŠธ๋งต์„ ๊ทธ๋ฆฌ๊ธฐ ์œ„ํ•ด ์›๋ณธ ์ด๋ฏธ์ง€์˜ ํ•ด์ƒ๋„๋ฅผ RR๋กœ ๋‚˜๋ˆ„์–ด ์ค„์ด๊ฒŒ ๋˜๋Š”๋ฐ(ํ•ด๋‹น ๊ทธ๋ฆฌ๋“œ์˜ ์ค‘์•™๊ฐ’์„ ์ค‘์‹ฌ ์ขŒํ‘œ๋กœ ์‚ฌ์šฉํ•œ๋‹ค), ์ด๋กœ ์ธํ•ด ๋ฐœ์ƒํ•˜๋Š” ์ •๋ณด์˜ ์†์‹ค์„ ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ๋‹ค. ์ด๋Š” ๊ทธ offset์— ๋Œ€ํ•œ L1 loss๋ฅผ cost์— ํฌํ•จํ•จ์œผ๋กœ์จ ๊ฐ€๋Šฅํ•˜๋‹ค. ์ด๋Ÿฌํ•œ offset loss๋Š” keypoint์— ํ•ด๋‹นํ•˜๋Š” ๊ทธ๋ฆฌ๋“œ์—๋งŒ ์ ์šฉํ•œ๋‹ค. ์ฆ‰ ๋‹ค๋ฅธ ๊ทธ๋ฆฌ๋“œ๋Š” maskingํ•˜์—ฌ loss๋ฅผ ๊ณ„์‚ฐํ•œ๋‹ค.
Loff=1Nโˆ‘pโˆฃO^p~โˆ’(pRโˆ’p~)โˆฃL_{off} = \frac{1}{N} \sum_p{| \hat{O}_{\tilde{p}} - (\frac{p}{R} - \tilde{p}) |}

Objects as Points

CenterNet์˜ ๊ฐ•์  ์ค‘ ํ•˜๋‚˜๋Š”, object detection๋ฟ ์•„๋‹ˆ๋ผ pose estimation, 3D detection ๋“ฑ ๋‹ค์–‘ํ•œ task์— ๋Œ€ํ•œ generalization์ด ์‰ฝ๋‹ค๋Š” ๊ฒƒ์ด๋‹ค. ์ด๋Š” ๊ธฐ๋ณธ์ ์œผ๋กœ ๊ด€๋ จ feature๋“ค์„ ์ด๋ฏธ์ง€์—์„œ ์ง์ ‘์ ์œผ๋กœ regressionํ•˜๊ธฐ์— ๊ฐ€๋Šฅํ•˜๋‹ค.
๋จผ์ €, object size๋ฅผ ์ค‘์‹ฌ์ ์— ํ•ด๋‹นํ•˜๋Š” ๊ทธ๋ฆฌ๋“œ์—์„œ ๋‹ค์Œ๊ณผ ๊ฐ™์€ L1 loss๋ฅผ ๊ณ„์‚ฐํ•จ์œผ๋กœ์จ ํ•™์Šตํ•œ๋‹ค.
Lsize=1Nโˆ‘kNโˆฃS^pkโˆ’skโˆฃL_{size} = \frac{1}{N} \sum_k^N{| \hat{S}_{p_k} - s_k |}
์ด์ œ detection์„ ์œ„ํ•œ loss๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๊ณ„์‚ฐ๋œ๋‹ค. ๋…ผ๋ฌธ์—์„œ๋Š” ฮปsize=0.1,ฮปoff=1\lambda_{size} = 0.1, \lambda_{off} = 1๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค.
Ldet=Lk+ฮปsizeLsize+ฮปoff+LoffL_{det} = L_k + \lambda_{size} L_{size} + \lambda_{off} + L_{off}
์ด์— ๋”ฐ๋ผ ๊ฐ ๊ทธ๋ฆฌ๋“œ์˜ ์ถœ๋ ฅ๊ฐ’์˜ ํฌ๊ธฐ๋Š” class๋ณ„ heat(C), size(2), offset(2)๋ฅผ ํฌํ•จํ•˜์—ฌ C+4C + 4๊ฐ€ ๋œ๋‹ค. ์—ฌ๊ธฐ์„œ ๊ฐ’์ด ๊ฐ€์žฅ ๋†’์€ 100๊ฐœ์˜ peaks๋“ค์„ ๋‚จ๊ธฐ๊ณ , ์ด๋“ค์— ๋Œ€ํ•ด ์œ„์˜ ์ •๋ณด๋“ค์„ ์ข…ํ•ฉํ•˜์—ฌ bounding box๋ฅผ ์ƒ์„ฑํ•จ์œผ๋กœ์จ detection์ด ๊ฐ€๋Šฅํ•ด์ง„๋‹ค.
๊ทธ๋Ÿฐ๋ฐ... ์—ฌ๊ธฐ์„œ NMS๊ฐ€ ๋”ฐ๋กœ ํ•„์š” ์—†๋‹ค๊ณ  ํ•˜๋Š”๋ฐ, ๊ทธ๋Ÿฌ๋ฉด ์ด๋ฏธ์ง€๋งˆ๋‹ค ๊ผญ 100๊ฐœ์˜ bounding box๋ฅผ ์ƒ์„ฑํ•˜๊ฒŒ ๋˜๋Š” ๊ฑด๊ฐ€...?

Human Pose Estimation

3D detection์€ ์–ด๋ ค์šด๋ฐ ์กฐ๋งŒ๊ฐ„ ์“ธ ์ผ์€ ์—†์„ ๊ฒƒ ๊ฐ™์œผ๋‹ˆ ์ƒ๋žตํ•˜๊ณ , CenterNet์„ pose estimation task๋กœ generalizeํ•˜๋Š” ๊ฒƒ์„ ์ •๋ฆฌํ•œ๋‹ค.
๊ธฐ๋ณธ์ ์œผ๋กœ human pose estimation๋Š” ์ธ๊ฐ„์˜ k=17k = 17๊ฐœ์˜ ๊ด€์ ˆ์„ ํŒŒ์•…ํ•จ์œผ๋กœ์จ ๊ฐ€๋Šฅํ•˜๋‹ค๊ณ  ์ „์ œํ•œ๋‹ค. ๋”ฐ๋ผ์„œ, CenterNet์—์„œ๋Š” pose estimation์„ ์œ„ํ•ด kร—2k \times 2 dimension์˜ "์ธ๊ฐ„ object์˜ ์ค‘์‹ฌ์ ์œผ๋กœ๋ถ€ํ„ฐ์˜ ๊ด€์ ˆ object์— ์ค‘์‹ฌ์ ๊นŒ์ง€์˜ offset" feature(JJ)๋ฅผ L1 loss๋ฅผ ํ†ตํ•ด ํ•™์Šตํ•œ๋‹ค.
์—ฌ๊ธฐ์„œ joint keypoint detection ๊ฒฐ๊ณผ๋ฅผ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด, kk class์˜ ๊ด€์ ˆ์— ๋Œ€ํ•œ keypoint heatmap์„ ์œ„์—์„œ ์„ค๋ช…ํ•œ ๋ฐฉ์‹๊ณผ ๋™์ผํ•˜๊ฒŒ ํ•™์Šตํ•œ๋‹ค. ์ฆ‰ Gaussian Kernel์„ ํ†ตํ•ด ํžˆํŠธ๋งต์„ ์ƒ์„ฑํ•˜๊ณ , ๊ฐ๊ฐ ๊ด€์ ˆ์˜ ์ค‘์‹ฌ์ ์— ๋Œ€ํ•œ offset์„ ํฌํ•จํ•˜์—ฌ ํ•™์Šตํ•œ๋‹ค.
๊ทธ๋ฆฌ๊ณ  ์ฒ˜์Œ์— ํ•™์Šตํ•œ ๊ด€์ ˆ offset J^\hat{J}์— ๋Œ€ํ•ด, ํ•ด๋‹น ๊ด€์ ˆ class์˜ ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด heatmap keypoint(confidence > 0.1์ธ peaks ๋“ค์„ ํ•„ํ„ฐ๋งํ•˜์—ฌ ๋ฝ‘์•„๋ƒ„) ์ขŒํ‘œ์„ ํ• ๋‹นํ•จ์œผ๋กœ์จ ์šฐ๋ฆฌ๊ฐ€ ์›ํ•˜๋Š” ๊ด€์ ˆ์˜ ์ขŒํ‘œ๋ฅผ ์ฐพ์•„๋‚ธ๋‹ค! ์ฆ‰ ์ง„์งœ ์ขŒํ‘œ๋Š” ํžˆํŠธ๋งต์œผ๋กœ ์ฐพ๋Š” ๊ฒƒ์ด๊ณ , ์ธ๊ฐ„์˜ ์ค‘์‹ฌ์ ๊ณผ ๊ทธ์— ๋Œ€ํ•œ ๊ฐ๊ฐ์˜ ๊ด€์ ˆ offset์€ grouping cue ์—ญํ• ์„ ํ•˜๋Š” ๊ฒƒ์ด๋‹ค. (ํ•ด์„์ด ํ‹€๋ ธ์„ ์ˆ˜ ์žˆ๋‹ค. ํ˜น์‹œ๋‚˜ ํ‹€๋ ธ๋‹ค๋ฉด ๋Œ“๊ธ€์ด๋‚˜ ์—ฐ๋ฝ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.)
์ด๋Ÿฌํ•œ CenterNet์˜ ์„ฑ๋Šฅ์€, ๊ฝค ๋น ๋ฅด๊ณ  ์ ๋‹นํ•œ ์ •ํ™•๋„๋ฅผ ๋ณด์—ฌ์ฃผ๋Š” ์ •๋„์ธ ๊ฒƒ ๊ฐ™๋‹ค. SOTA ๊ธ‰์˜ ์ •ํ™•๋„๋Š” ๋‹น์—ฐํžˆ ์•„๋‹ˆ๊ณ , ๊ทธ๋ ‡๋‹ค๊ณ  ๋ง‰ ๋ฏธ์นœ๋“ฏ์ด ๋น ๋ฅธ ๊ฒƒ๋„ ์•„๋‹ˆ์ง€๋งŒ, ๊ทธ ๋ฐธ๋Ÿฐ์Šค๋ฅผ ์ž˜ ๋งž์ถ˜ ์ƒํƒœ์—์„œ ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ์„ ์ž๋ž‘ํ•˜๋Š” CenterNet(?)!
E.O.D.