REAL-TIME LIGHT FALL-OFF STEREO Miao Liao, Liang Wang, Ruigang Yang University of Kentucky {mliao3,lwangd,ryang}@cs.uky.edu Minglun Gong Memorial University of Newfoundland gongml@gmail.com ABSTRACT We present a real-time depth recovery system using Light Fall-off Stereo (LFS). Our system contains two co-axial point light sources (LEDs) synchronized with a video camera. The video camera captures the scene under these two LEDs in complementary states(e.g., one on, one off). Based on the in- verse square law for light intensity, the depth can be directly solved using the pixel ratio from two consecutive frames. We demonstrate the effectiveness of our approach with a number of real world scenes. Quantitative evaluation shows that our system compares favorably to other commercial real-time 3D range sensors, particularly in textured areas. We believe our system offers a low-cost high-resolution alternative for depth sensing under controlled lighting. 1. INTRODUCTION Many applications, such as robot navigation and augmented reality, require real-time range information in a dynamic envi- ronment. In this paper we developed a novel system that uses the inverse square law for light intensity to estimate depth in- formation. Based on the formulation in [1] our system uses a single camera to capture a scene under two different lighting conditions: one illuminated by a near point light source and the other by a far one. Per-pixel depth is solved based on the pixel intensity ratio and the distance between the two lights, without the need for matching pixels. The main contribution of this paper is a novel depth range system that can generate a VGA (640 × 480) resolution depth map at 30Hz. Quantitative accuracy evaluation shows that our system compares favorably to other commercial 3D range sensors, particularly in textured areas. In addition, our system is made of commodity off-the-shelf components, offering an inexpensive solution to real-time, high-resolution, video-rate range sensing. 1.1. Related work Recovering 3D shapes from images is one of the fundamental tasks in computer vision. While there is a plethora of tech- niques to achieve this, we will focus on the methods that are capable of generating real-time depth maps with live input. The most common way of computing depth map is to use stereovision. Recently, several stereo methods have been de- veloped to exploit the processing power of modern graphics hardware [2, 3, 4, 5]. Although tremendous progress has been made in stereovision, the fundamental correspondences prob- lem remains difficult in real-world applications. The correspondence problem can be greatly simplified with active illumination. Many real-time structured light scanners (e.g. [6, 7, 8]) can obtain high quality results. These systems typically require multiple frames, which limit the ob- ject motion, and have difficulty with high-frequency textures. New range sensors have also been developed using shut- tered light-pulse (SLP) technologies [9]. 3DV Systems, Ltd. and Canesta, Inc. [10, 11] have both developed SLP technolo- gies. However They are either very expensive (e.g. over fifty thousand US dollars for a 3DV system) or have limited reso- lutions (e.g., 64 × 64 for a Canesta sensor). Our system builds on the algorithms described in [1] which use the inverse-square law to recover 3D shape infor- mation. Compared to previously developed techniques, our approach only requires two images and the use of commod- ity off-the-shelf components provides an inexpensive way to produce high-resolution depth maps. More importantly, experiments show that our system provides better depth maps that are independent of scene texture. 2. METHODS 2.1. Light Fall-off Stereo It is well known that the intensity of light emitted from a source of constant intrinsic luminosity falls off as the square of the distance from the object. Under this inverse square law, the observed intensity of a surface point p can be formulated as: I p = L(θ) r 2 p ρ(θ,φ), (1) where L(θ) is the light radiance along incident direction θ. r p is the distance between the light source and p. ρ(θ,φ) is the BRDF (Bidirectional Reflectance Distribution Function) of surface point p and φ is the viewing direction.