Lighting Model-Guided Initialization for Gradient Descent-Based Projector Compensation

Woochan Sun , Geonu Noh , Kwanghee Ko , Bochang Moon

Gwangju Institute of Science and Technology

IEEE Access

Lighting Model-Guided Initialization for Gradient Descent-Based Projector Compensation
Projection mapping results of the gradient descent-based projector compensation framework, PCDR [16], on a non-planar curved surface with textured colors (a). PCDR adjusts the projector input images ((b) and (c)) so that the projected outputs ((e) and (f)) within the target display area (denoted by a cyan box in (a)) match the user-provided target image (d). We compute an initial guess of the projector input using a lighting model and pass this to PCDR, allowing it to begin its iterative optimization from our output rather than its default initialization (i.e., a constant input). This modification enables PCDR to produce visually and numerically improved results without altering its underlying algorithm.

Abstract

Projection mapping, which maps a projector input (i.e., an image) onto a physical surface, is widely used to display user-specified visual content. However, the projected content can become distorted when an ideal surface, such as a flat and white plane, is unavailable. As a result, adjusting the projector input using a projector compensation technique becomes necessary to ensure that the projected image matches the user-provided target image. While various techniques have been proposed for projector compensation, a recent approach models projection mapping as a simulatable process in virtual space, where a light transport algorithm simulates the projection mapping. In such rendering frameworks, projector compensation is typically achieved by iteratively adjusting the projector input via a gradient descent-based optimizer, starting from an initial guess (often chosen arbitrarily). In this paper, we investigate how to set the initial values more effectively rather than choosing them arbitrarily. As the main contribution of the paper, we propose a new initialization scheme that determines the starting values based on a lighting model. We then integrate this model-guided initialization into gradient descent-based optimization and demonstrate that it improves projection mapping results, particularly for non-planar and colored surfaces.

Contents