Artacho Bruno ; Savakis Andreas

Abstract

We propose BAPose, a novel bottom-up approach that achieves state-of-the-artresults for multi-person pose estimation. Our end-to-end trainable frameworkleverages a disentangled multi-scale waterfall architecture and incorporatesadaptive convolutions to infer keypoints more precisely in crowded scenes withocclusions. The multi-scale representations, obtained by the disentangledwaterfall module in BAPose, leverage the efficiency of progressive filtering inthe cascade architecture, while maintaining multi-scale fields-of-viewcomparable to spatial pyramid configurations. Our results on the challengingCOCO and CrowdPose datasets demonstrate that BAPose is an efficient and robustframework for multi-person pose estimation, achieving significant improvementson state-of-the-art accuracy.