Paper Reading on Pose Detection

Simple baselines for Human Pose Estimation and Tracking

• Pose Estimation Using A deconvolution Head network
  • obtaining high resolution feature maps is crucial, but no matter how
• Pose Tracking
  • human detector
  • optical flow
  • greedy matching
  • flow-based pose similarity metric
• Joint Propagation Using optical flow
  • motion blur
  • $(x+\delta x, y + \delta y)$
• Flow-based Pose similarity
  • $S_{Flow}(J_i^k,J_j^l) = OKS(\hat{J}_i^l, J_j^l)$

Lightweight

Mobilenet

• depthwise separable convolutions
  • a layer for filtering, a layer for combining
  • $\displaystyle{\frac{D_K \cdot D_K\cdot M\cdot D_F\cdot D_F + M \cdot N\cdot D_F\cdot D_F}{D_K \cdot D_K\cdot M\cdot N\cdot D_F\cdot D_F}=\frac{1}{N}+\frac{1}{D_K^2}}$
• Width multiplier & resolution multiplier

Mobilenet v2

• inverted residual with linear bottleneck

Mobiblenet v3

MCUNet