2020-456 Passive 3D Shape via Polarization-Aware Neural Networks

SUMMARY

UCLA researchers in the Department of Electrical Engineering and Computer Science have developed an algorithm to take image data and generate accurate 3D shape representations using a deep neural network framework. The method uses data from polarization imaging to generate 3D models that are less sensitive to noise and that are five times more accurate than pure physics-based algorithms.

BACKGROUND

Accurate 3D models help designers and end users visualize space requirements and improve efficiency and accuracy without having to create a physical model. Simple applications of this include quality control, where the shape of complex parts are measured to ensure accuracy. More complex applications include automation where robotics need accurate models in order to navigate and manipulate their surroundings. While methods to generate 3D models from measured data exist, they either require physical measurement through touch (which is slow and is impossible for certain materials) or non-invasive measurement (which use complex physics to convert the image data into the models and have a lot of noise due to variances in lighting and illumination of the object). There is a need for a fast and reliable method that combines the computational efficiency of physics algorithms with the robust nature of machine learning to generate 3D models that are less sensitive to noise.

INNOVATION

Researchers have developed a hybrid algorithm that uses both physics equations as well as data driven machine learning to generate three dimensional models that are five times more accurate then pure physics-based algorithms. The method takes data (e.g. image data) and expected outcome (e.g. 3D shape) and uses deep neural networks to learn on their own how to predict shape from image data. The advantage of this is that the machine can learn to ignore aspects like variation in lighting or refractive surfaces to generate more accurate predictions.

POTENTIAL APPLICATIONS