Luminance Ratios for Stable SIFT Keypoint Detection in Challenging Illuminations Scenarios

Invariance to changes in scene illumination is a desirable feature in computer vision, but its accomplishment is still a challenge even for modern systems. This paper proposes a novel way to detect keypoints in scale-space by exploring the concept of luminance ratio first introduced in Wallach's Lightness Constancy Theory, instead of using linear difference operators. The hypothesis is that an operator based on the concept of luminance ratio may provide more robust results with respect to illumination changes. In order to test this hypothesis, the SIFT keypoint detector was adapted to the proposed approach with minimum implementation effort. Experimental results show that the luminance ratio approach indeed yields more stable keypoint detection than the original SIFT based on linear difference operators, specially in lower illumination conditions. Existing SIFT implementations can be easily adapted to use the luminance ratio concept for improved operation in environments with uncontrolled lighting conditions.