85th European Study Group with Industry

16th–20th April 2012, University of East Anglia, Norwich

Determining 3D Body Shape from 2D Images

A problem brought to the 85th European Study Group with Industry by Poikos Ltd..

Problem Description

The overall aim is to use two digital photographs of a person to compute the person’s 3-D shape. The photographs are taken with ‘ordinary’ equipment; a monocular lens in a smartphone, or a web-camera in a lap-top computer would be typical. There is no specification on the brand of equipment.

The individual is requested to wear close-fitting clothing and to adopt particular poses for the two photographs, at a distance of about 3m from the camera. One image is face-on to the camera, and the other is side-on to the camera. Poikos ask the person his/her height for a sense of scale.

Poikos would like to invite the Study Group to challenge their existing approach by investigating methods that combine the segmentation and 3D reconstruction (stereology) into a single process. In addition, Poikos would also be interested to learn of methods that can be used to reduce errors associated with camera defects. Furthermore, Poikos would like to hear ideas for means of determining scale without the need for standardised objects to be present in the image.

Study Group Report

The study group addressed two particular issues in the overall process that Poikos would like to improve on, a markerless correction for radial distortion and improved segmentation of the person's outline from the image. Based on a radial distortion function the study group deduced and implemented an optimization method for Finding the function parameters given straight lines in the distorted image. For the segmentation problem, the study group applied Perona–Malik pre-processing to improve edge detection in the image. An open source version of the 'segmentation by weighted aggregation' method was applied to the images and shows considerable promise. Together with prior information of the content of the image, this algorithm could provide better results than the current Poikos segmentation method.