Detect pose landmarks within bounding boxes

Use This When

• Building pose estimation systems for fitness tracking, gesture recognition, or action analysis
• Extracting skeletal keypoints for PPE detection where body part localization is needed
• Creating human-computer interfaces that respond to body language or hand gestures
• Preparing landmark data for downstream orientation normalization or body part extraction

What It Does

• Detects pose landmarks (face, body, hands) within provided bounding boxes using Triton-served models
• Returns nested landmark lists with x,y coordinates and confidence for each person/region
• Supports multiple pose estimation models including SimCC-based and heatmap-based approaches
• Handles multi-person scenarios by processing each bounding box independently

Works Best With

• Person detector → this component → orient-landmarks or extract-landmarks for normalization
• Body pose → calculate-landmark-angle for gesture classification or posture analysis
• Face detector → this component → facial landmark visualization or expression analysis
• Feeding detect-bodypart to isolate head/hands/torso regions for PPE classification

Caveats

• Requires accurate bounding boxes; poor detection quality degrades landmark precision
• Landmark schema (count, order, semantic meaning) varies by model; verify compatibility
• SimCC models output different format than heatmap models; configure output parsing correctly
• Small or heavily occluded subjects yield low-confidence landmarks that may need filtering