Spring 2016 Seminar Series

DATE: 3/31/16
LOCATION: E51-325
TIME: 4:15pm
Reception immediately following

SPEAKER:
Sebastian Pokutta

TITLE
High Accuracy Body Part Learning from 3D Depth Images

ABSTRACT
Learning body parts, body shape, joint locations, and more generally anthropometric measures from visual data is an important task in computer vision and machine learning as it forms the basis for many human-computer interfaces. Usually these learning tasks are performed either on 2D RGB images and/or 3D depth images and there is a plethora of algorithms capable of performing this task. These algorithms are optimized for real-time performance, motion tracking, and stability across a huge variety of body poses and clothing and in consequence often trade-off estimation accuracy. In healthcare and medical diagnostics applications however the requirements are very different. Measurement and estimation accuracy trumps real-time speed and motion tracking requirements. Moreover, the patient is often directed into a specific pose so that pose invariance is secondary as well. As such the performance trade-off of many of the aforementioned algorithms is suboptimal for medical diagnostics. We present a two-pass machine learning approach to identify body parts and estimate anthropometric measures with very high accuracy. We present two variants, one based on random forests and another based on deep learning. Training is performed with the diagnostics applications in mind via a customized CGI pipeline, where a large variety of synthetic human body models is generated simulating anthropometric feature progression (e.g., various stages of a medical condition) and final validation is performed on real-world data. We will also discuss applications of the obtained algorithm in ongoing real-world projects.

 

Joint work with A. Roy and D. Zink