Ballester Irene ; Peterka Ondřej ; Kampel Martin

Abstract

3D Human Pose Estimation (HPE) is the task of locating keypoints of the humanbody in 3D space from 2D or 3D representations such as RGB images, depth mapsor point clouds. Current HPE methods from depth and point clouds predominantlyrely on single-frame estimation and do not exploit temporal information fromsequences. This paper presents SPiKE, a novel approach to 3D HPE using pointcloud sequences. Unlike existing methods that process frames of a sequenceindependently, SPiKE leverages temporal context by adopting a Transformerarchitecture to encode spatio-temporal relationships between points across thesequence. By partitioning the point cloud into local volumes and using spatialfeature extraction via point spatial convolution, SPiKE ensures efficientprocessing by the Transformer while preserving spatial integrity per timestamp.Experiments on the ITOP benchmark for 3D HPE show that SPiKE reaches 89.19%mAP, achieving state-of-the-art performance with significantly lower inferencetimes. Extensive ablations further validate the effectiveness of sequenceexploitation and our algorithmic choices. Code and models are available at:https://github.com/iballester/SPiKE