[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): SSIR(xi,xj)=||f(xi)−f(xj)||22
  • CIR: SCIR(xi,xj)=wTg(xi,xj)−b
• Also combine pairwise loss and triplet loss.
  
  • Pairwise
    LPSIR=∑i,j[1+hij(||f(xi)−f(xj)||22−bSIR)]+
    LPCIR=αP2||w||22+∑i,j[1+hij(wTg(xi,xj)−bSIR)]+
    where bSIR,bCIR is the distance threshold (margin), αP is the trade-off parameter, which is set to 0.0005 in the experiments
    Combine: LP=LPSIR+ηPLPCIR
  • Triplet
    LTSIR=∑i,j,k[bSIR−||f(xi)−f(xk)||22+||f(xi)−f(xj)||22)]+
    LTCIR=αP2||w||22+∑i,j,k[bCIR+wTg(xi,xk)−wTg(xi,xj)]+
    Combine: LT=LTSIR+ηTLTCIR
• Compute cross-image feature map
  φr(xi,xj)=max(0,br+∑qkq,r∗ϕq(xi)+lq,r∗ϕq(xj))

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.