Deep Network for Integrated 3D Sensing of Multiple People in Natural Images

Explore a deep network approach for integrated 3D sensing of multiple people in natural images in this 22-minute video lecture from the University of Central Florida. Delve into the problem of human localization and grouping, learning about the objective to find skeletal joints and the use of Panoptic Studio for ground truth. Examine the computational pipeline, including deep volume encoding, limb scoring, and 3D pose decoding with shape estimation. Discover the SMPL (Skinned Multi-Person Linear Model) and its application in deep autoencoders. Analyze the learning process and review both quantitative and qualitative results using the CMU Panoptic dataset, gaining insights into advanced computer vision techniques for multi-person 3D sensing in natural environments.

Problem: Human Localization and Grouping
Objective
Approach: Find the Skeletal Joints
Ground Truth: Panoptic Studio
Computational Pipeline
Deep Volume Encoding (2d/3d)
Limb Scoring
3d Pose Decoding and Shape Estimati
Given the feature volume of a person and it's skeleton
SMPL: A Skinned Multi-Person Linear Model
Deep Autoencoder
Analysis of learning
Quantitative Results
CMU Panoptic dataset
Qualitative Results