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).