Replacing Goniophotometer with Camera and U-Net with Hypercolumn Rescale Block

We deal with replacing a costly and slow goniophotometer device with a standard, inexpensive, and fast camera in the task of evaluating an illuminated area by a car headlamp. This solution is novel, has not yet been solved, and has the potential to speed up the process of prototyping headlamps. The difficulties lie in the significantly different resolutions of the two devices and in the disparity between intensities captured by the camera and goniophotometer due to the nonlinear behavior of the light. We propose to capture images by a camera with various exposure times and handle them as a multispectral image. The image is processed by U-Net architecture where we replaced the standard decoder with a Hypercolumn rescale block. The proposed scheme produces a mean absolute percentage difference between the real goniophotometer and our solution of less than 0.5%.