AI BriefingLG AI Research: 3D Human Pose and Shape Estimation Research
Key point
LG AI Research has proposed a framework that efficiently handles complex spatiotemporal relationships in video to precisely estimate 3D human pose and shape.
Details
Existing single-frame-based 3D human pose and shape estimation methods suffer from unstable performance in situations involving motion blur or occlusion. To address this, existing video-based methods have used Global Average Pooling to compress spatial information while modeling spatiotemporal relationships, but this process caused reconstruction errors.
This research proposes a new framework that considers both spatial and temporal dimensions while efficiently managing complexity. The key technologies are as follows:
- Spatial Alignment Module (SAM): To resolve spatial discrepancies between adjacent frames, it uses Affine Transformation to align adjacent features to the central feature.
- Space2Batch: Decomposes spatial and temporal relationships to dramatically reduce the complexity of spatio-temporal attention.
- Uncertainty-guided Attention Re-weighting: Enhances the model's robustness even in challenging environments such as motion blur or complex backgrounds.
This model uses the SMPL parametric model to estimate shape and pose parameters, and achieved SOTA (State-of-the-Art) performance on major benchmark datasets, with detailed results presented at WACV 2024.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
MultiMatte, a Promptable Background Removal Model Based on SAM 3, Released, Achieving S-measure of 0.901 on DIS-VD Benchmark