上海科技大学知识管理系统

Children's HMR is a critical task with significant applications, yet current methods fall short of effectively addressing its challenges. Existing approaches, primarily developed for adults, struggle when applied to pediatric subjects due to fundamental differences in anatomical ratios, distinct pose distributions, and the limited availability of dedicated training data for children. This limitation arises from an overreliance on adult-centric body models and datasets that fail to capture the nuanced morphological and kinematic variations of younger individuals. In this work, we present a novel approach that tackles these challenges by leveraging mesh-aligned evidence during training to capture detailed structural information and keypoint-aligned evidence during inference to refine mesh recovery. This dual-evidence strategy improves the model's generalization, particularly on out-of-distribution data. Moreover, by integrating advanced 3D understanding with the semantic capabilities of Large Language Models (LLMs), our framework achieves enhanced robustness. We validate our method on the AGORA dataset and present comparative results on several in-the-wild images, demonstrating its superior performance and resilience compared to existing techniques. Overall, our work not only narrows the gap between adult-centric and child-specific HMR but also lays a promising foundation for further advancements in this domain.