Reconstructing car shape from a still image

This paper addresses the problem of detecting and recognizing car 3D-contour in the presence of car occlusion from a still image. We proposed a deformable template—Meaning Active Basis Model (mABM) which is a variant of Active Basis Model (ABM). MABMs consists of a small number of Gabor wavelet elements with different locations, orientations and sizes, and every element corresponds some specific part contour of a car. We use mABMs to detect 2D-contours of cars with various colors, sizes and shapes, in the process of detection elements of mABM are allowed to slightly perturb their locations and orientations to get the accurate 2D-contours. After detection, the car 3D shape is reconstructed according to the 2D-contour and the calibrated camera parameters. Another contribution of this paper is a novel technique of handling car occlusion: we coarsely rebuild car 3D-cuboid models according to the detecting result of mABMs, followed by the estimation of the overlap coefficients to separate occluded cars. Experimental results with real world images show that this method is effective in the detection of cars even under severe occlusions.