(CVPR2020)Learning a Neural Solver for Multiple Object Tracking

Learning a Neural Solver for Multiple Object Tracking

主要贡献

基于Message Passing Networks (MPNs)提出一种全可微的框架
直接在图域上操作，使得本方法可以在整个检测集上直接作出推论。

By operating directly on the graph domain, our method can reason globally over an entire set of detections and predict final solutions.
We propose a novel time-aware neural message passing update step inspired by classic graph formulations of MOT.
网络对MOT的学习不仅仅局限于特征提取，而是可以深入到数据关联过程中

we show that learning in MOT does not need to be restricted to feature extraction, but it can also be applied to the data association step.

简介

tracking-by-detection 分为目标检测和数据关联两步。目标检测有很多基于深度学习的检测器。数据关联往往被视为图分割问题。从图的观点来看MOT，节点代表着检测到的目标，边则意味着俩个节点有所关联。一条激活的边两端的节点属于同一个追踪轨迹。解决图分割问题也可以分为两步，首先计算每条边的cost，cost反映了两端的节点属于同一个轨迹的可能性。随后cost被用来获得最终优化后的分割图。

两个方向：优化图的表达，优化cost的选择与计算。

(i) learn features for MOT, and (ii) learn to provide a solution by reasoning over the entire graph.

并不成对的计算cost然后计算匹配，本文通过整张图直接预测最终的轨迹。

message passing network (MPN) 直接在图域上学习。

因此，我们的方法能够解释探测之间的全局相互作用，尽管依赖于一个简单的图表公式。

Hence, our method is able to account for global interactions among detections despite relying on a simple graph formulation.

We show that our framework yields substantial improvements with respect to state of the art, without requiring heavily engineered features and being up to one order of magnitude faster than some traditional graph partitioning methods.