Computer Vision News - September 2019

7 Results One of the most problematic aspects of IID is the difficulty to quantify the results. It is very hard to gather direct ground truth data for this task. Previous works have collected ground truth via painting object, synthetic rendering and manual annotation, but each of these methods has a significant limitation. Here we show you some qualitative results, so that you can examine by yourself the quality of the suggested method. Below is a figure that demonstrates the results during the optimization iterations. Each column is the reflectance image (lower) and the shading image (upper) for each iteration 1, 3, 5, 7. We can see, just like the method intended, the reflectance component of the decomposition becoming flattened/sparse while the shading becoming smoother. This demonstrates the convergence of the iterative method. In the next figure we can see the final results of the IID task. There is a separation of shadows and illumination from the light source to shading component and color consistency in the reflectance component. These are quite good results. As we can see enforcing the suggested priors does seem to generate good quality decomposition. Intrinsic Image Decomposition (IID)