Previous work on variable gain image sequences

Fully automated computational approaches to simultaneously estimate:

• the unknown nonlinear response function of the camera;
• the unknown overall change in gain caused by automatic exposure control or AGC; and
• the projective coordinate transformation relating the images to one another,

were presented (implemented and shown) in [2], and further explored in [3] using both parametric and nonparametric methods. A goal of this work was to combine variable gain image sequences into a single image of increased spatiotonal range and definition, as well as for a front end to a wearable vision system. More recently, Szeliski also considered the problem of estimating only the projective coordinate transformation between images [4], while Debevec and Malik [5] have considered the problem of estimating only the camera response function. Mitsunaga and Nayar have also considered the problem of estimating the response function using a low order polynomial [6]. Mann has also considered parametric estimates of the camera response function, by proposing a simple three parameter function that provides a very good fit to most camera response functions [7].

This paper concentrates on nonparametric determination of camera response functions. The problem of nonparametric reverse engineering a camera's response function, from differently exposed images of identical or overlapping subject matter, was first proposed and first solved in [2]. In this paper we present, in detail, such a computationally efficient maximum likelihood estimation based on least squares.