mmNorm Offers Peek Performance

Wonderful Wizards (that may also be charlatans) can no longer be sure that their true identities will remain a secret. Unfortunately for them, curtains (and walls, boxes, and plastic containers) can no longer hide what is behind them, thanks to the work of a group at MIT. They have developed a technology that is able to peer through all sorts of obstacles to reveal what is behind them, much like Superman’s X-ray vision.

This technology also has applications outside of Oz and Metropolis. It could be used in a warehouse, for instance, to look inside a sealed box to determine what it contains, and if the contents are in good shape. Applications in security are also ideal for this system — it could be used at airports to enable more accurate reconstructions of objects inside suitcases.

The researchers call their system mmNorm, and it leverages millimeter wave (mmWave) signals, like the ones used by Wi-Fi, to produce 3D reconstructions of objects that are hidden from view. While these signals pass through surfaces like interior walls and boxes, they reflect off of others. Through a novel approach, the team has found a way to analyze the reflections to recognize even small objects and fine details that past approaches could not resolve.

Traditionally, radar systems that use mmWave signals rely on a method called backprojection. This technique works well when identifying large objects, like vehicles obscured by fog, but fails when trying to discern smaller items such as utensils or tools. These failures stem from a limitation in resolution and an inability to accurately interpret the way surfaces reflect mmWave signals.

MIT’s mmNorm sidesteps this problem by using an entirely different strategy. Instead of focusing solely on where the signals bounce back from, the system estimates something called a surface normal, or the direction that a surface is facing at any given point. By collecting information from multiple radar locations and analyzing the strength of the reflections received, mmNorm effectively votes on what direction a surface is facing. When compiled, this voting process produces a highly accurate map of the object's curvature.

To test their idea, the researchers mounted a mmWave radar on a robotic arm that moves around a target, gathering reflections from many angles. A unique mathematical model, borrowed in part from computer graphics, then converts these reflections into a detailed 3D reconstruction of the hidden object.

In a series of tests, mmNorm achieved a reconstruction accuracy of 96% on over 60 real-world objects, compared to 78% for the best prior methods. It also excelled at detecting fine details like the curves of a mug handle or the differences between a knife and a fork buried together under clutter.

The project’s code and data have been released publicly, and the researchers plan to continue refining the system, improving its resolution, and extending its capabilities through denser materials. But already, mmNorm has significantly improved our ability to see what was once hidden.