Scholarlink

<p>3D human pose estimation (HPE) predicts 3D joint coordinates from 2D images or videos. Despite advances via deep learning, current methods often overlook textual information and human knowledge, which can provide valuable implicit supervision. To address these limitations, we introduce a novel approach called the improved Fine-Grained Prompt-Driven Denoiser (FGPD) with Temporal Constriction and Proliferation (TCP) transformer, built on a diffusion model. FGPD includes three key elements: (1) the Fine-grained Part-aware Prompt (FPP) module, which creates detailed part-aware prompts by integrating accessible textual data and domain-specific knowledge about body parts with learnable prompts to provide implicit guidance; (2) the Fine-grained Prompt-Pose Communication (FPC) module enables detailed interaction between part-aware prompts and pose data, improving denoising; and (3) the Prompt-driven Timestamp Stylization (PTS) module, which combines learned prompt embeddings with temporal noise level information to enable adaptive adjustments during each step of the denoising process. The Refined Temporal Constriction and Proliferation Transformer (RTCPT) combines spatio-temporal encoders with a TCP framework to capture multi-scale attention and address depth ambiguity. It includes a Feature Aggregation Refinement (FAR) module using a cross-layer strategy within the TCP block. Substantial experiments on the Human3.6M and MPI-INF-3DHP datasets demonstrate that our approach achieves superior performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Exploring multi-transformer with fine-grained prompt-driven coupled with diffusion model for 3D human pose estimation

  • 期刊论文

  • 发表日期：

    2026年2月9日

查看全文
  • Sathiyamoorthi Arthanari,
  • Sathishkumar Moorthy,
  • Jae Hoon Jeong

摘要

3D human pose estimation (HPE) predicts 3D joint coordinates from 2D images or videos. Despite advances via deep learning, current methods often overlook textual information and human knowledge, which can provide valuable implicit supervision. To address these limitations, we introduce a novel approach called the improved Fine-Grained Prompt-Driven Denoiser (FGPD) with Temporal Constriction and Proliferation (TCP) transformer, built on a diffusion model. FGPD includes three key elements: (1) the Fine-grained Part-aware Prompt (FPP) module, which creates detailed part-aware prompts by integrating accessible textual data and domain-specific knowledge about body parts with learnable prompts to provide implicit guidance; (2) the Fine-grained Prompt-Pose Communication (FPC) module enables detailed interaction between part-aware prompts and pose data, improving denoising; and (3) the Prompt-driven Timestamp Stylization (PTS) module, which combines learned prompt embeddings with temporal noise level information to enable adaptive adjustments during each step of the denoising process. The Refined Temporal Constriction and Proliferation Transformer (RTCPT) combines spatio-temporal encoders with a TCP framework to capture multi-scale attention and address depth ambiguity. It includes a Feature Aggregation Refinement (FAR) module using a cross-layer strategy within the TCP block. Substantial experiments on the Human3.6M and MPI-INF-3DHP datasets demonstrate that our approach achieves superior performance.