3D Human pose estimation from video via multi-scale multi-level spatial temporal features

In this paper, we present an innovative framework for 2D-to-3D human pose estimation from video, harnessing the power of multi-scale multi-level spatial-temporal features. Our framework comprises three integral branch networks: A temporal feature core network, dedicated to extracting temporal coherence among frames, enabling a comprehensive understanding of dynamic human motion. A multi-scale feature branch network, equipped with multiple receptive fields of varying sizes, facilitating the extraction of multi-scale features, thus capturing fine-grained details across different scales. A multi-level feature branch network, tasked with extracting features from layers at various depths within the architecture, providing a nuanced understanding of pose-related information. Within our framework, these diverse features are seamlessly integrated to encapsulate intricate spatial and temporal relationships inherent to the human body. This integration effectively addresses challenges such as depth ambiguity and self-occlusions, culminating in substantially improved accuracy in pose estimation.Extensive experiments on Human3.6M and HumanEva-I show that our framework achieves competitive performance on 2D-to-3D human pose estimation in video. Code is available at: https://github.com/fll123/3Dhumanpose.