Novel view synthesis for traffic accident scene from sparse uncalibrated sample views

Abstract

Based on limited sample views of a road traffic accident scene captured on site, this paper aims to achieve goal of virtually capturing the traffic accident scene from several virtual viewpoints by generating novel views of scene at novel viewpoints. To avoid explicit 3D reconstruction, the novel view is synthesized via a view interpolation mechanism, in which the to-be-interpolated image function changing with viewpoint is formulated to 2 sub-functions: image feature position and image intensity. The synthesized novel view changes smoothly with the novel viewpoint in terms of both feature position and intensity and returns a given sample view exactly when the novel viewpoint falls upon the viewpoint associated with this sample view. Experimental result shows that, even from sparse uncalibrated sample views, the proposed method is effective in producing novel views that appear like those captured by a slowly moving camera.