Freeform pixels are camera sensing elements of arbitrary shape, each built from a photodetector behind an optical mask whose continuous transmittance performs a linear projection of the scene. Because an array of such pixels is mathematically a bias-free linear layer, the camera optics and the downstream inference network are trained together as a single neural network, and the learned first-layer weights become the fabricated mask patterns. This lets a camera solve lightweight tasks such as counting people or estimating traffic speed with only a handful of measurements, which in turn preserves privacy and enables fully self-powered operation.
A minimalist vision system uses the fewest pixels needed for a task by replacing a camera's fixed square-pixel grid with freeform pixels: photodetectors behind optical masks of arbitrary continuous t…