Enabling Effective Low-Light Perception using Ubiquitous Low-Cost Visible-Light Cameras

Abstract

Deep ConvNets have changed and improved the capabilities and robustness of machine perception models such as object detectors. Meanwhile, visual perception using visible-light sensors under low-light conditions is still challenging for deep learning models. While there are sensors that allow mitigating some of the limitations of visible-light cameras, they are often infeasible in real-life cost-, power- or space-constrained systems. By contrast, to extend the functionality of existing vision systems based on visible-light cameras, we investigate effective machine perception in low light using these ubiquitous low-cost sensors. Our first contribution to the computer vision community was our high-quality object detection dataset targeting low-light machine perception. Next, we proposed a neural Image Signal Processor that processes raw sensor data into representation optimal for low-light machine cognition. In the future work, we aim to focus on a higher-level approach to low-light machine cognition. Finally, we plan to address low-light conditions holistically by integrating low- and high-level domain solutions in a common framework.