Learning to reason from uncalibrated wide angle images

点击翻译按钮获取中文摘要

We witness an increasing demand of using cameras with wide field of view (FOV) lenses in many real-life applications, including security, augmented reality (AR), healthcare and autonomous systems. The increased field of view minimizes cost, energy and computation since fewer cameras are needed. The downside of using such lenses is that they require the use of lens calibration methods to obtain a lens distortion profile, which is then used to rectify the image-that is, to cancel the effects of distortion and create a perspective projection image. A wide array of such methods, ranging from classical (deterministic) to deep-learning-based have been proposed. Unfortunately, calibration is considered as a burden in most critical systems due to the many pre- and post-processing steps.In this proposal, we propose to break free from the "calibrate-and-rectify" paradigm, and instead directly reason on the wide-angle images. The goal of this project is therefore to develop a set of theoretical and experimental tools for scene understanding on images captured with uncalibrated, wide-angle lenses.