Approximated Slack Scaling for Structural Support Vector Machines in Scene Depth Analysis

Algorithm 2

The parameter learning for slack method.

Input: data 𝑋_𝑡, label 𝑌_𝑡, size T, tolerance 𝜀

Initialize parameter 𝑊 → 0, constraint set 𝑅 → ∅

Repeat

  for t = 1 to T

    𝑌^∗=argmin_𝑌[𝑠(𝑦)+𝜉/Δ(𝑦_𝑡,𝑦)]

  end for

  increase constraint set 𝑅←𝑅∪{𝑌^∗}

  (𝑊,𝜉)← solve the QP using all the existing 𝑌^∗

Until ∑_𝑡[𝜉/Δ(𝑦_𝑡,𝑦)+𝑠(𝑦_𝑡)−𝑠(𝑦)]≤𝜉+𝜀