Pixel-based correspondence and shape reconstruction for moving objects

We present a method for pixel-based correspondence and shape reconstruction from multiple views of a moving object. Instead of relying on the brightness constancy assumption, this correspondence measure explicitly models intensity variations among multiple views, thus allowing estimation of lighting conditions and surface normals. The frontal-parallel assumption is not required either, since only a single intensity from each view is required for this pixel-based correspondence. The proposed shape reconstruction algorithm enforces lighting consistency and produces a final surface by fusing triangulated 3D positions and estimated surface normals without imposing smoothness constraints. Both synthetic and real experiments are conducted, where the surfaces of textureless and textured objects can be successfully reconstructed.