Multiple People Tracking by Lifted Multicut and Person Re-identification

CVPR2017, MOTS

Posted by pshow on June 30, 2020

Multiple People Tracking by Lifted Multicut and Person Re-identification

主要贡献

  • we propose a novel graph based formulation that links and clusters person hypotheses over time by solving an instance of a minimum cost lifted multicut problem.
  • ReID模型包含了外观与姿势特征。
  • propose to cast multi-person tracking as the minimum cost lifted multicut problem. We introduce two types of edges (regular and lifted edges) for the tracking graph.

简介

使用CNN提取的特征有很强的鲁棒性,也有利于ReID

外观相似的目标并不一定属于同一个ID

similar looking people are considered as the same person only if they are connected by at least one feasible track (possibly skipping occlusion)

每一个检测框都被表述为图中的一个节点。带有损失权重的边将跨越时间,空间将节点连接起来。

优点:

  • 追踪目标的个数没有理论定义上的限制或者偏重。
  • 由于同一帧中同一个目标的重复检测会被自然的聚类在一起,所以可以取消掉启发式的NMS算法

In order to avoid that distinct but similar looking people are assigned to the same track, a distinction must be made between the edges that define possible connections (i.e., a feasible set) and the edges that define the costs or rewards for assigning the incident nodes to distinct tracks (i.e., an objective function).

相关工作

流问题,最小损失分割问题。

模型

边的权重被描述为两个节点属于同一ID的可能性

image-20200630110728961

Parameters.

image-20200630111637394

常规的边会连接同一帧的节点或者低于阈值的附近帧的节点。

lifted edges 提升边链接时间距离较远的相似目标。

Feasible Set.

图分割问题

image-20200630115006307

image-20200630115609785

image-20200630135937798

Objective function.

image-20200630140840693

边的可能性越小,边的cost越大。

Optimization.

image-20200630141255871

Person Re-identification for Tracking

Architectures

image-20200630142957550

image-20200630143940354

Fusing Body Part Information

image-20200630144051124

image-20200630144258118

image-20200630144514264

ReID实验分析

Pairwise Potentials

image-20200630145620664

image-20200630145959108

image-20200630150401961

Tracking Experiments and Results

image-20200630153143885

image-20200630153909722

主要探讨$\delta_t$ 和$\delta_{max}$对MP , LMP 的影响。


本文浏览量: