Improving object recognition with the ℓ-channel

We augment a new channel called the ¿-channel to conventional RGB images, and propose its application in multiple classification and recognition tasks. The new RGB-¿ image records the same scene using the color and frosted light channel, which are simultaneously captured using a binocular camera with a low-cost frosted glass placed in front of one of the cameras. Due to the light scattering property of the frosted glass, the acquired frosted light channel is imprecise. In this paper we propose a novel optimization that is guided by the RGB channel to refine the ¿-channel to preserve edges due to scene radiance. Extensive experimental results have demonstrated the effectiveness of our RGB-¿ images, where significant improvements are reported in a variety of scene classification and object recognition tasks. HighlightsWe propose a new image format RGB-l by augmenting a new channel called ¿ -channel to RGB images.The acquisition cost of the new channel is very cheap using a frosted glass.Our additional channel is refined in a novel framework solving a unified optimization problem.Extensive experimental results justify the effectiveness of our novel lightness channel for many recognition problems.