The Inference & 3D Reconstruction Pipeline subsystem orchestrates the complete process from raw input to final 3D human pose and shape results, serving as the primary user API.
Components
ROMP Inference API
Primary API for ROMP inference. Orchestrates the entire inference pipeline for the ROMP model, including model loading, executing the forward pass, and initiating model-specific post-processing.
Referenced Source Code
BEV Inference API
Primary API for BEV inference. Orchestrates the entire inference pipeline for the BEV model, including model loading, executing the forward pass, and initiating model-specific post-processing.
Referenced Source Code
TRACE Inference API
Primary API for TRACE inference. Orchestrates the entire inference pipeline for the TRACE model, including model loading, executing the forward pass, and initiating model-specific post-processing.
Referenced Source Code
Core Inference Engine
Provides the fundamental prediction logic, preparing neural network modules and executing the forward pass to produce raw model outputs. This component is shared across different model APIs.
Referenced Source Code
Center Map Processor
Generates and parses 2D/3D center maps from the raw model outputs to identify and localize human instances, serving as an initial step in post-processing.
Referenced Source Code
General Result Parser
Processes the initial parsed outputs, matches predicted parameters to instances, and suppresses duplicate detections to refine results before model-specific adjustments.
Referenced Source Code
Model-Specific Post-processing
Performs specialized post-processing steps tailored to the unique outputs and requirements of their respective models (ROMP, BEV, TRACE), refining the general parsed results into a format suitable for 3D reconstruction.
Referenced Source Code
3D Reconstruction (SMPL)
Generates 3D human meshes and keypoints based on the processed SMPL parameters, forming the final output of the pipeline.