Project

Calibration Invariant Imaging

Guy Satat, Matthew Tancik, Otkrist Gupta, Barmak Heshmat, Ramesh Raskar

Groups

Object Classification through Scattering Media with Deep Learning


A method for classifying objects hidden behind a scattering layer with a neural network. Training on synthetic data with variations in calibration parameters allows the network to learn a model that doesn't require calibration during lab experiments.

Traditional techniques to see through scattering media rely on a physical model that needs to be precisely calibrated. Computationally overcoming the scattering relies heavily on accurately calibrated physical models. Thus, such systems are extremely sensitive to a precise and lengthy calibration process.

In this work we overcome this bottleneck by utilizing neural networks and their ability to learn models that are invariant to data transformation. In our case, the transformations are variations in the imaging system calibration parameters. To that end, we create a synthetic dataset that contains variations in all calibration parameters (we use a Monte Carlo forward model to render the measurements). The system is then tested on actual lab experiments without specific calibration or tuning.

Object Classification through Scattering Media with Deep Learning


A method for classifying objects hidden behind a scattering layer with a neural network. Training on synthetic data with variations in calibration parameters allows the network to learn a model that doesn't require calibration during lab experiments.

Traditional techniques to see through scattering media rely on a physical model that needs to be precisely calibrated. Computationally overcoming the scattering relies heavily on accurately calibrated physical models. Thus, such systems are extremely sensitive to a precise and lengthy calibration process.

In this work we overcome this bottleneck by utilizing neural networks and their ability to learn models that are invariant to data transformation. In our case, the transformations are variations in the imaging system calibration parameters. To that end, we create a synthetic dataset that contains variations in all calibration parameters (we use a Monte Carlo forward model to render the measurements). The system is then tested on actual lab experiments without specific calibration or tuning.