Material appearance is traditionally represented through its Bidirectional Reflectance Distribution Function (BRDF), quantifying how incident light is scattered from a surface over the hemisphere. To speed up the measurement process of the BRDF for a given material, which can necessitate millions of measurement directions, image-based setups are often used for their ability to parallelize the acquisition process: each pixel of the camera gives one unique configuration of measurement. With highly specular materials, the High Dynamic Range (HDR) imaging techniques are used to acquire the whole BRDF dynamic range, which can reach more than 10 orders of magnitude. Unfortunately, HDR can introduce star-burst patterns around highlights arising from the diffraction by the camera aperture. Therefore, while trying to keep track on uncertainties throughout the measurement process, one has to be careful to include this underlying diffraction convolution kernel. A purposely developed algorithm is used to remove most part of the pixels polluted by diffraction, which increase the measurement quality of specular materials, at the cost of discarding an important amount of BRDF configurations (up to 90% with specular materials). Finally, our setup succeed to reach a 1.5° median accuracy (considering all the possible geometrical configurations), with a repeatability from 1.6% for the most diffuse materials to 5.5% for the most specular ones. Our new database, with their quantified uncertainties, will be helpful for comparing the quality and accuracy of the different experimental setups and for designing new image-based BRDF measurement devices.