Colour Transfer By Feature Based Histogram Registration

A common problem in computer vision is that different sensors acquire different colour responses to an imaged object. This problem occurs because physical factors during the imaging process introduce a variation that differs for each sensor; in addition, it is practically impossible to image an object under perfectly constant lighting conditions at different spatial positions within an imaging environment. This variation degrades the performance of colour computer vision processes such as object tracking; in addition, the involved nature of calibration routines means that the calibration step is often ignored.

The colour transfer approach offers the potential for automatic alignment of similar colour spaces without manual intervention to perform the calibration. The goal of colour transfer methods is to make the image regions for the same object the same colour; in this paper, we seek to achieve this by aligning the corresponding dense regions of the histograms (clusters). We are only concerned with the situation when the images contain the same or at least highly similar objects. Existing colour transfer methods assume that scale changes between corresponding clusters are small; however, a significant change in object scale is common when the camera moves or when tracking between multiple cameras.

The authors developed a scale invariant cluster alignment algorithm and pose colour transfer as a two-frame registration problem where their goal is to align the corresponding clusters in two colour histograms in the presence of lighting, automatic camera setting and object scale variation. They have tested the algorithm for aligning a range of colour histograms obtained from pairs of images with systematically introduced variation of lighting, scale and shadowing. The experiments have shown quantitative improvements in colour histogram alignment over the simple methods of aligning moments. Authors have shown the merits of a feature-based histogram registration approach for aligning multi-modal data and how to deal with its limitations.