Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Open posedoc

7,962 views

Published on

This slide is expolanation about OpenPose Model which realtime recognise poses of multiple people.

Published in: Data & Analytics
  • Be the first to comment

  • Be the first to like this

Open posedoc

  1. 1. OpenPose Realtime Multi-Person 2D Pose Estimation using Part Anity Fileds By Zhe Cao @The Robotics Institute CMU mabonki0725 ()11 June 30, 2017
  2. 2. Real Time 2017/04/14 https://arxiv.org/abs/1611.08050 OpenPose GitHub PBL@AIIT openPose OpenPose Real Time 20Frame/sec) Figure: OpenPose 2 / 10
  3. 3. OpenPose Real Time Top-down ottom-up Top-down OpenPose Bottom-up ? ! @ A 3 / 10
  4. 4. OpenPose Real Time Bottom-up ( CNN F CNN S1 F CNN L1 ! T F St−1 Lt−1 CNN St F St−1 Lt−1 CNN Lt PAF ) Figure: CNN 4 / 10
  5. 5. CNN CNN ft S = J j=1 p W(p)||St j(p) − S∗ j (p)||2 2 ft L = J j=1 p W(p)||Lt j(p) − L∗ j (p)||2 2 L = T t=1 ft S + ft L ! j P St j t j Lt j t j S∗ j j L∗ j j 5 / 10
  6. 6. CNN j S∗ j,k = exp − ||P − xj,k||2 2 σ2 S∗ j = max k S∗ j,k(P) # k k xj,k k j σ Figure: ) 6 / 10
  7. 7. PAF(Partial Anity Filed) Ec vk = xj2,k − xj1,k ||xj2,k − xj1,k||2 $ L∗ c,k(p) = vk vk vk⊥|p − xj1,k| σl 0 (otherwise) % L∗ c (p) = 1 nc(p) k L∗ c,k(p) Ec = u=1 u=0 Lc(p(u)) dj2 − dj1 ||dj2 − dj1 ||2 du ' k k xj,k k j c Figure: PAF 7 / 10
  8. 8. PAF(Partial Anity Filed) 2 z max z Ec = max Zc m∈Dj1 n∈Dj2 Emnzmn j1,j2 s.t n∈Dj2 zm,n j1,j2 ≤ 1 m∈Dj1 zm,n j1,j2 ≤ 1 Ec zm,n j1,j2 8 / 10
  9. 9. PAF(Partial Anity Filed) n 2 n 2 NP Hard ? ! @ A Figure: n 9 / 10
  10. 10. OpenPose Realtime Figure: 10 / 10

×