search

AbnerHqC / GaitSet

A flexible, effective and fast cross-view gait recognition network
582 stars 170 forks source link

Seperate Triplet Loss #120

Open lonelygoatherd opened 4 years ago

lonelygoatherd commented 4 years ago

请问在FramePipeline中直接将(n,m,d)输入seperate triplet loss，为何是seperate triplet loss，其seperate是每个scale的特征1个triplet loss吗？为何不直接像testing时拉直整个feature为(batch_size, hidden_dim)进而用1个“普通”的triplet loss？