Part?MOT: A multi?object tracking method with instance part?based embedding

Part-MOT, a one-stage anchor-free architecture which unifies the object identification representation and detection in one task for visual tracking is presented. For representation, position relevant feature obtained using center-ness information, takes advantage of ideal to encode map as instance-aware embedding. To adapt object's movement, clustering-based method get global instance introduced. This enables this approach more robust make better decisions. Part-MOT achieves state-of-the-art performance on public datasets, with especially strong results deformation movement changes.