Towards multi-person pose tracking: Bottom-up and top-down methods
ICCV PoseTrack Workshop(2017)
摘要
In this paper, we focus on the challenging problem of multi-person pose tracking in the wild. Recent multi-person articulated tracking methods can be categorized into the top-down and bottom-up approaches. We investigate the advantages and disadvantages of both bottom-up and topdown methods on various datasets. We propose a novel bottom-up joint detector, termed as MSPAF to extract multiscale features and implement a human detector based on recent development of object detection. Incorporating the global context, we use a human detector to rule out bottomup false alarms which significantly improves the tracking results. Following the commonly used graph partitioning formulation, we construct a spatio-temporal graph and solve a minimum cost multicut problem for human pose tracking. Our proposed method achieves the state-of-the-art performance on both” Multi-Person PoseTrack” dataset and” ICCV 2017 PoseTrack Challenge” dataset.
更多查看译文
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络