CVPR 2022 - Tesla AI proposed: generalized pedestrian re recognition based on graph sampling depth metric learning

In recent days, , Dr. Shaoling, chief scientist of Tesla technology group, and his team proposed an efficient small batch sampling method （mini-batch sampling） Method —— Figure sampling （Graph Sampling, GS）, For large-scale in-depth measurement learning , It greatly improves the generalization of pedestrian recognition . at present , The research results （ Titled : Graph Sampling Based Deep Metric Learning for Generalizable Person Re-Identification） This year's CVPR Accepted and published .

Pedestrian re recognition is a popular computer vision task , Its goal is to retrieve a large number of image base , So as to find the pedestrians in the given query image . In the past two years , Generalized pedestrian re recognition has attracted more and more attention because of its research and practical value . This kind of research explores the generalization of learning pedestrian recognition model for unprecedented scenes , And it uses direct cross data set evaluation to conduct performance benchmark .

At present, the popular methods of deep learning pedestrian recognition models include classification （ Use ID loss）、 Measure learning （ Use pairwise loss or triplet loss）, And their combination （ for example ID + triplet loss）.ID The loss function is very convenient for classification learning . However , In large-scale deep learning , Involving classifier parameters will generate a lot of memory and computing costs in the forward and back propagation process . Similarly , It is also inefficient to involve the category related parameters used to measure learning in the global view .

chart 1： Two different sampling methods ：（ left ）PK Sampler ;（ On the right side ） Proposed by Dr. Shaoling's team GS Sampler . Different shapes represent different categories , Different colors represent different batches （batches）.GS Build a diagram for all categories , And always sample the nearest adjacent category