Jaewook Kang
Nov. 23th 2018
1
© 2018
MoT Labs
All Rights Reserved
J. Kang Ph.D. presents
2
J. Kang Ph.D. presents
Bio.
(2018, 10-~)
최근에는 자연어 처리도 쬐금씩...
3
Jaewook Kang (강재욱)
누구나 TensorFlow!
J. Kang Ph.D.
Proj. Contributors
4
Jaewook Kang
(Principal 삽질러)
Taekmin Kim
(해커톤 중독자)
Doyoung Kwak (iOS, coreml 갓)
Jeongah Shin
(Android tflite 갓)
DongSeok Yang
(야망개발자2)
Yonggeun Lee
(야망 개발자1)
J. Kang Ph.D. presents
오늘의 이야기
5
단순한 논문 구현은 그만! 프로젝트를 시작하자!�- 거북목 프로젝트의 시작과 딥러닝 캠프�
���
세상에 가치를 주는 일을 하자!�
6
딥러닝 역량 성장법
7
딥러닝 역량 성장법
→ 딥러닝을 위한 기본 지식 습득
→ 딥러닝 최신 동향 + 딥러닝 도구 사용법 숙달
→ 딥러닝을 통한 문제 해결 경험
8
거북목 프로젝트의 시작
논문 그만 보고
세상에 가치를 주는 프로젝트를 하자!
9
거북목 프로젝트의 시작
10
Project Outline
11
Approach Sketch
12
Estimate the coordinate of four body parts:
Approach Sketch
13
14
일단 놀고
낮에 미친듯이 일하고
밤에 잠안자고 놀고 (민폐)
주말에 놀고
개발하느라 힘들었던 한달
(사실 두달)
19
6월
7월
Frameworks:
- Tensorflow
Jaewook Kang et al.
20
Don’t Be Turtle Proj
Prototype App v0.4 (18.Aug)
J. Kang Ph.D. presents
Frameworks:
- Tensorflow
Jaewook Kang et al.
21
Don’t Be Turtle Proj
Prototype App v0.4 (18.Aug)
J. Kang Ph.D. presents
앱은 봄이 오기전에 마켓에 나올꺼예요!
22
J. Kang Ph.D. presents
Pose Estimation 10min Tour
for 거북목 프로젝트
거북목 프로젝트도 결국 RnD
23
© 2017-2018
Jaewook Kang
All Rights Reserved
누구나 TensorFlow!
J. Kang Ph.D.
프로젝트의 핵심
24
저현고 인공지능 특강
J. Kang Ph.D.
프로젝트의 핵심: Pose Estimation!
25
저현고 인공지능 특강
J. Kang Ph.D.
Human Pose Estimation
26
누구나 TensorFlow!
J. Kang Ph.D.
Human Pose Estimation
27
누구나 TensorFlow!
J. Kang Ph.D.
2D single pose estimation
28
누구나 TensorFlow!
J. Kang Ph.D.
Human Pose Estimation
Localization + Classification
29
누구나 TensorFlow!
J. Kang Ph.D.
Human Pose Estimation
Localization + Classification
30
(x,y)=(0.34,0.92)
(x,y)=(0.34,0.92)
(x,y)=(0.34,0.92)
(x,y)=(0.34,0.92)
누구나 TensorFlow!
J. Kang Ph.D.
Human Pose Estimation
Localization + Classification
31
Head
Neck
Rshoulder
Lshoulder
누구나 TensorFlow!
J. Kang Ph.D.
Human Pose Estimation
32
Pose coordinate
Prediction
누구나 TensorFlow!
J. Kang Ph.D.
DeepPose (Alexander’14)
33
누구나 TensorFlow!
J. Kang Ph.D.
DeepPose (Alexander’14)
34
누구나 TensorFlow!
J. Kang Ph.D.
DeepPose (Alexander’14)
35
Image credit: http://www.1000ventures.com/design_elements/selfmade/elephant_holistic-6perceptions.png
누구나 TensorFlow!
J. Kang Ph.D.
DeepPose (Alexander’14)
36
누구나 TensorFlow!
J. Kang Ph.D.
DeepPose (Alexander’14)
37
누구나 TensorFlow!
J. Kang Ph.D.
DeepPose (Alexander’14)
38
누구나 TensorFlow!
J. Kang Ph.D.
DeepPose (Alexander’14)
Toshev, A., Szegedy, C., “Deeppose: Human pose estimation via deep neural networks,” CVPR 2014
39
누구나 TensorFlow!
J. Kang Ph.D.
DeepPose (Alexander’14)
40
누구나 TensorFlow!
J. Kang Ph.D.
DeepPose (Alexander’14)
41
누구나 TensorFlow!
J. Kang Ph.D.
DeepPose (Alexander’14)
42
: 회귀 함수
: 전처리 함수
: 후처리 함수
누구나 TensorFlow!
J. Kang Ph.D.
DeepPose (Alexander’14)
43
AlexNet
X
Input image
Coordinate prediction
Head = (0.1,0.3)
Neck = (0.2,0.6)
Rshoulder = (0.8,0.1)
….
Y
True coordinate
(y0,y1,....,yk)
L2 loss
Train
.
.
.
.
전처리
후처리
누구나 TensorFlow!
J. Kang Ph.D.
DeepPose (Alexander’14)
Global context !
즉 신체 구조에 대한 이해 부족!�
Special thank to 피카소! :-)
누구나 TensorFlow!
J. Kang Ph.D.
DeepPose (Alexander’14)
�
누구나 TensorFlow!
J. Kang Ph.D.
DeepPose (Alexander’14)
�
.
*
1st regression
(y0,y1)^S=0
2nd regression
(y0,y1)^S=1
누구나 TensorFlow!
J. Kang Ph.D.
DeepPose (Alexander’14)
AlexNet
X with i-th box
from prev stage
Displacement
prediction
L2 loss
Train
Displacement
from prev stage
.
Single stage regressor
전처리
후처리
누구나 TensorFlow!
J. Kang Ph.D.
DeepPose (Alexander’14)
X
Input image
True coordinate
Y=(y0,y1,....,yk)
1st stage
regressor
.
S-th stage
regressor
Coordinate
prediction
Head = (0.1,0.3)
Neck = (0.2,0.6)
Rshoulder = (0.8,0.1)
….
.
.
X
Input image
Displacement
누구나 TensorFlow!
J. Kang Ph.D.
DeepPose (Alexander’14)
�
49
누구나 TensorFlow!
J. Kang Ph.D.
Convolutional Heatmap Regressor
�
50
누구나 TensorFlow!
J. Kang Ph.D.
Convolutional Heatmap Regressor
51
image credit: https://arxiv.org/abs/1609.01743
누구나 TensorFlow!
J. Kang Ph.D.
Convolutional Heatmap Regressor
52
Model
X
Input image
Heatmap prediction
Y
True coordinate
(y0,y1,....,yk)
Some loss fn
Train
Heatmap
generator
.
.
.
True
Heatmap
.
누구나 TensorFlow!
J. Kang Ph.D.
Convolutional Heatmap Regressor
53
누구나 TensorFlow!
J. Kang Ph.D.
Multiscale Understanding
54
누구나 TensorFlow!
J. Kang Ph.D.
Multiscale Understanding
55
누구나 TensorFlow!
J. Kang Ph.D.
Multiscale Understanding
56
누구나 TensorFlow!
J. Kang Ph.D.
Multiscale Understanding
57
image credit: https://arxiv.org/abs/1602.00134
누구나 TensorFlow!
J. Kang Ph.D.
Multiscale Understanding
58
누구나 TensorFlow!
J. Kang Ph.D.
Multiscale Understanding
59
누구나 TensorFlow!
J. Kang Ph.D.
Beyond...
60
누구나 TensorFlow!
J. Kang Ph.D.
모바일 앱에 올리기 위한 모델사이즈 줄이기�
���어디 한번 실전에 적용해 보자!���
61
J. Kang Ph.D. presents
Our Baseline Model
62
J. Kang Ph.D. presents
Our Baseline Model
63
J. Kang Ph.D. presents
Our Baseline Model
64
J. Kang Ph.D. presents
65
사이즈를 줄이기 위한 방향
66
J. Kang Ph.D. presents
사이즈를 줄이기 위한 방향
67
J. Kang Ph.D. presents
사이즈를 줄이기 위한 방향
68
J. Kang Ph.D. presents
사이즈를 줄이기 위한 방향
�
69
J. Kang Ph.D. presents
거북목앱에 Global Context이해가 필요해?
70
1/4배
Proposed: Single HG model
J. Kang Ph.D. presents
거북목 앱에 Multi-scale 이해가 필요해?
71
Conventional: four stages
J. Kang Ph.D. presents
거북목 앱에 Multi-scale 이해가 필요해?
72
Proposed: two stages
J. Kang Ph.D. presents
Step1
Step2
Step3
Step4
855.8MB
220.3MB
258.4MB
4.56MB
1.Baseline
2. Single HG model
+ Inverted bottleneck
3. Feature Space
Optimization
5. HG Stage
Reduction
X0.3
X0.006
1.55MB
3.83MB
2.17MB
<1MB ??
거북목 프로젝트 모델 줄인 이야기
J. Kang Ph.D. presents
Sample Prediction
�
74
Batchsize=64, lr=1e-3,
Alpha=0.0625
Expansion rate =10
2.19MB
Batchsize=64, lr=1e-3,
Alpha=0.0625,
Expansion rate =7
1.55MB
J. Kang Ph.D. presents
Sample Prediction
�
75
Batchsize=64, lr=1e-3,
Alpha=0.0625
Expansion rate =10
2.19MB
Batchsize=64, lr=1e-3,
Alpha=0.0625,
Expansion rate =7
1.55MB
J. Kang Ph.D. presents
Sample Prediction
�
76
Batchsize=64, lr=1e-3,
Alpha=0.0625
Expansion rate =10
2.19MB
Batchsize=64, lr=1e-3,
Alpha=0.0625,
Expansion rate =7
1.55MB
J. Kang Ph.D. presents
Sample Prediction
�
77
Batchsize=64, lr=1e-3,
Alpha=0.0625
Expansion rate =10
2.19MB
Batchsize=64, lr=1e-3,
Alpha=0.0625,
Expansion rate =7
1.55MB
J. Kang Ph.D. presents
오늘의 이야기
78
79
야호! 끝났다!
이제 DevFest를 즐기자!