Neural Network Generalization: The impact of camera parameters
This work addresses the practical problem of neural network robustness to camera variations for computer vision applications, though it is incremental in analyzing specific parameters.
The study quantified how camera parameters affect convolutional neural network generalization for car identification, finding that generalization between different cameras is similar to that between camera and synthetic images, and that pixel size, demosaicking, and bit-depth significantly impact performance.
We quantify the generalization of a convolutional neural network (CNN) trained to identify cars. First, we perform a series of experiments to train the network using one image dataset - either synthetic or from a camera - and then test on a different image dataset. We show that generalization between images obtained with different cameras is roughly the same as generalization between images from a camera and ray-traced multispectral synthetic images. Second, we use ISETAuto, a soft prototyping tool that creates ray-traced multispectral simulations of camera images, to simulate sensor images with a range of pixel sizes, color filters, acquisition and post-acquisition processing. These experiments reveal how variations in specific camera parameters and image processing operations impact CNN generalization. We find that (a) pixel size impacts generalization, (b) demosaicking substantially impacts performance and generalization for shallow (8-bit) bit-depths but not deeper ones (10-bit), and (c) the network performs well using raw (not demosaicked) sensor data for 10-bit pixels.