Exemplar-Based Colour Constancy

Abstract

We address the problem of scene depth recovery within cross-spectral stereo imagery (each image sensed over a differing spectral range). We compare several robust matching techniques which are able to capture local similarities between the structure of crossspectral images and a range of stereo optimisation techniques for the computation of valid depth estimates in this case. Specifically we deal with the recovery of dense depth information from thermal (far infrared spectrum) and optical (visible spectrum) image pairs where large differences in the characteristics of image pairs make this task significantly more challenging than the common stereo case. We show that the use of dense gradient features, based on Histograms of Oriented Gradient (HOG) descriptors, for pixel matching in combination with a strong match optimisation approach can produce largely valid, yet coarse, dense depth estimates suitable for object localisation or environment navigation. The proposed solution is compared and shown to work favourably against prior approaches based on using Mutual Information (MI) or Local Self-Similarity (LSS) descriptors.

A Hybrid Strategy for Illuminant Estimation Targeting Hard Images

“…It is interesting to note that there are some images considered hard for the statistical-based method that all learning-based method are successful on. Overall, however, the L1 (exemplar-based [23]) method does particularly well for the hard images, able to produce a better result on all except a few of the images.…”

“…Statistical-based methods including: S1 = shades of grey [13], S2 = grey world [4], S3 = 1 st order grey edge [27], S4 = 2 nd order grey edge and S5 = white-patch [25]. Learning-based methods including: L1 = exemplar-based [23], L2 = color constancy using natural image statistics [16], L3 = edgebased gamut, L4 = pixel-based gamut, L5 = intersectionbased gamut [14,18], L6 = Bayesian method [15] and L7 = spatial correlation [5].…”

“…The centroid (mean) of the five estimates is calculated and used with the SVM to predict if the image is hard or easy. If the image is classified as hard, we use a learning-based method such as the exemplar-based method [23] to process the image to obtain the final illumination estimate.…”

“…Their performance, however, is generally not as good as learning-based methods. Learning-based methods (representative examples include [14,12,16,15,18,5,11,23]) exploit the availability of training images that have labelled ground truth illumination. These methods use image features to train regressors to predict the illumination based on the input image and associated training-data.…”

“…In the case an image is categorised as hard, it is likely that the results of the simple camera on-board white balancing algorithms are incorrect. Such hard images can be saved as RAW on the camera for later off-line processing by slower, but more accurate, learning-based methods, such as the exemplar-based method [23,24]. This leads to better overall illumination estimation performance while reducing the overall time-complexity.…”