Revealed at the recent SIGGRAPH 2016 in Anaheim, California, the new Body Talk system is the innovative work of a team of researchers from the Max Planck Institute for Intelligent Systems and the University of Texas at Dallas. Researchers at the Planck Institute’s Perceiving Systems Department – which is where the system was developed – work to “seek mathematical and computational models that formalize the principles of vision.” As a result, Body Talk was developed as way to create accurate 3D body models from 2D images using multiple linguistic descriptions of body shapes.
From mental representations to 3D models
“We hypothesized that people have a shared understanding of shape, and that this is reflected in our use of language. If we agree on what shape means, then we conjecture that the collective judgements of shape attributes by ‘the crowd’ contains a robust signal about body shape,” says one of the team’s researchers about the basis of the system. Using suggestions from 15 volunteers, who rate body shape using 30 words or fewer, the Body Talk system is able to produce a metrically accurate 3D avatar simply from descriptions. The team sees the technology having useful capabilities in a range of industries including online retail, gaming, virtual reality and healthcare sectors. “When scanning technology is unavailable or inapplicable, Body Talk can be used to visualize mental representations of human body shape,” said Dr. Michael J. Black, Director of the Perceiving Systems Department at the Max Planck Institute for Intelligent Systems. The system’s potential in the place of scanning technologies is endless because of its ability to predict images from conceptual ideas and arbitrary symbols, as opposed to physical objects.
More than just words
The above video demonstrates the full development and usability of the Body Talk system. The surprising accuracy of Body Talk’s 3D representations proves that adjectives and common phrases regarding body shape are better and clearer indicators for visual concepts than people’s own perceptions of themselves. One researcher on the project noted that, “remarkably, words alone predict linear measurements with an error below one centimeter on average, and circumferences to about 1.4 centimeters. We can use our model to better understand what shape words mean. The shape descriptors are correlated and our model captures those correlations…The resulted body shapes correspond to our prototypical, shared idea of human shape.” Ultimately, the final product is a combination of traditional meanings which people have come to agree upon. Therefore, Body Talk is not yet able to accurately represent specific, abnormal body attributes such as those belonging to celebrities or other widely recognized figures, but offers reliable representations of average, commonly occurring body shapes and types. To learn more about Body Talk and its application, please visit here.
