Revisiting and evaluating colour constancy and colour stabilisation algorithms

When we capture a scene with a digital camera, the sensor generates a digital response which is the Raw image. This response depends on the ambient light, the object reflectance and the sensitivity of the camera. The generated image is processed with the the camera pipeline, which is a series of operations aiming at processing the colours of the image to make it more pleasant for the user. Further colour processing can also be performed on the pipeline output image. This said, processing the colours is not only important for aesthetic reasons, but also for various computer vision tasks where a faithful reproduction of the scene colours is needed e.g. for object recognition and tracking. In this thesis, we focus on two important colour processing operations: colour constancy and colour stabilisation. Colour constancy is the ability of a visual system to see an object with the same colour independently of the light colour; the camera processes the image so the scene looks like captured under a canonical light, usually a white light. This means that when we take two images of, let’s say, a green apple in the sunlight and indoor under a tungsten light, we want the apple to appear green in both cases. To do that one important step of the pipeline is to estimate the light colour in the scene to then discount it from the image. In this thesis we first focus on the illuminant estimation problem, in particular on the performance evaluation of illuminant estimation algorithms on the benchmark ColorChecker dataset. More precisely, we show the importance of the accuracy of the ground-truth illuminants when evaluating algorithms and comparing them. The following part of the thesis is about chromagenic illuminant estimation which is based on using two images of the scene: one filtered and one unfiltered where the two images need to be registered. We revisit the preprocessing step (colour correction) of the chromagenic method and we introduce the use of the Monge-Kantorovitch transform (MKT) that removes the need for the expensive registration task. We also introduce two new datasets of chromagenic images for the evaluation of illuminant estimation methods. The last part of the thesis is about colour stabilisation which is particularly important in video processing, where consistency of colours is required across image frames. When the camera moves or when the shooting parameters change, the same object in the scene can appear with different colours in two consecutive frames. To solve for colour stabilisation given a pair of images of the same scene we need to process the first image to match the second. We propose using MKT to find the mapping. Our novel method gives competitive results compared to other recent methods while being less computationally expensive