LightTrack: A Generic Framework for Online Top-Down Human Pose - PowerPoint PPT Presentation

LightTrack: A Generic Framework for Online Top-Down Human Pose Tracking Authors: Guanghan Ning, Heng Huang

LightTrack: Online Multi-Person Pose Tracking q Overview of Proposed Framework 2 6/1/20 1:54 PM

Proposed Framework q Single-Person Pose Tracking Incorporate object tracking with single-person pose tracking. A recursive approach: • Use human keypoints as explicit visual cues for bounding box tracking. • Estimate keypoints for future frames with tracked bounding boxes. • Detections only performed at Key-frames. 3 6/1/20 2:09 PM

Proposed Framework q Re-ID with Pose Matching Pose Matching Network: Siamese Graph Convolutional Network (SGCN) 4 6/1/20 2:17 PM

Proposed Framework q Re-ID with Pose Matching • In some situations, poses are indeed similar. • Spatial Consistency is helpful as well. 5 6/1/20 2:22 PM

Performance q Qualitative Results 6 6/1/20 1:54 PM

Performance q Qualitative Results Unseen Test Video from PoseTrack Dataset Unseen Surveillance Video 7 6/1/20 2:32 PM

Performance q Quantitative Results 8 6/1/20 2:32 PM

Performance q Benchmark for Multi-person Pose Tracking PoseTrack’17 9 6/1/20 2:32 PM