PFSeq: Photorealistic Fisheye Sequence

A photorealistic fisheye sequence to evaluate self-calibration, stereo matching, feature points detection and matching, sky segmentation. This simulated sequence has been created with Blender (using Cycles engine and the Compositor) between 2013 and 2015. It depicts an urban canyon, with fisheye cameras facing the sky. It has been motivated by the absence of previous similar data, and by the need of perfect ground truth, that remains an open problem with real acquisition. To reach photorealistic rendering, care has been done on following aspects: Buildings’ textures are issued from real photos. A blueish area makes the transition of the image circle *. Purple fringing issued from chromatic aberrations *. Lens flare when the sun is visible *. Optical dispersion blur *. Motion blur *. Cameras have a high resolution of 1200×1200 pixels and an effective field of view of 181.8 degrees. Two variants with different fisheye projections: equidistant projection: r = f . θ equisolid angle projection: r = 2 . f . sin(θ/2) For each variant, data are available in two versions: with and without realistic optical and motion distortions, depicted with a *. More information on PFSeq are available at my personal page.