[Paper note] Joint Learning of Single-image and Cross-image Representations for Person Re-id.

Intuition

• Re-id method: single-image representation (SIR) and cross-image representation (CIR).
• Combine them together!

Method

• SIR measurements are special cases of CIR-based classification.
  
  • SIR(Euclidean distance): $S_{SIR}(x_i,x_j)=||f(x_i)-f(x_j)||_2^2$
  • CIR: $S_{CIR}(x_i,x_j)=w^Tg(x_i,x_j)-b$
• Also combine pairwise loss and triplet loss.
  
  • Pairwise
    $L_{SIR}^P=\sum_{i,j}[1+h_{ij}(||f(x_i)-f(x_j)||_2^2-b_{SIR})]_+$
    $L_{CIR}^P=\frac{\alpha_P}{2}||w||_2^2+\sum_{i,j}[1+h_{ij}(w^Tg(x_i,x_j)-b_{SIR})]_+$
    where $b_{SIR},b_{CIR}$ is the distance threshold (margin), $\alpha_P$ is the trade-off parameter, which is set to 0.0005 in the experiments
    Combine: $L^P=L_{SIR}^P+\eta_PL_{CIR}^P$
  • Triplet
    $L_{SIR}^T=\sum_{i,j,k}[b_{SIR}-||f(x_i)-f(x_k)||_2^2+||f(x_i)-f(x_j)||_2^2)]_+$
    $L_{CIR}^T=\frac{\alpha_P}{2}||w||_2^2+\sum_{i,j,k}[b_{CIR}+w^Tg(x_i,x_k)-w^Tg(x_i,x_j)]_+$
    Combine: $L^T=L_{SIR}^T+\eta_TL_{CIR}^T$
• Compute cross-image feature map
  $\varphi_r(x_i,x_j)=max(0,b_r+\sum_qk_{q,r}*\phi_q(x_i)+l_{q,r}*\phi_q(x_j))$

Result

• Dataset: CUHK-01/03, VIPeR.
• 52.17% on CUHK03.
• Investigation on sensitivity of trade-off parameter.
• 71.8% on CUHK01 (pretrain on CUHK03).
• 35.76% on VIPeR.