Exploring Histogram-Based Color Constancy

David R. Treadwell IV, Yunxuan Rao, Daniel Y. Bi, Bruce Allen Maxwell

Abstract: We present a lightweight fully convolutional network for color constancy (LHCC). The network uses multiple 2-D projections of the 3-D log RGB histogram of an image in order to predict the color correction coefficients. In developing the network, we explored whether to use linear RGB or log RGB data, the network structure (width and depth), how to handle dark pixels, how to generate the 2-D mappings of the 3-D histogram, and how to normalize or transform the bin counts in order to preserve the fine histogram structure. Our results show that attention to each of these details makes a difference in overall performance. Our most significant findings are that using log RGB outperforms linear RGB for this task, and that using a log transformation of the bin counts outperforms thresholding, a hyperbolic tangent, and linear normalization. Our exploration resulted in a fully convolutional network with 0.5M parameters that sets a new performance standard on SimpleCube++ with a surprising 2.72◦angular error on the worst 25%. It is also competitive on older data sets such as Gehler-Shi and NUS-8.