Registration and Matching of Perspective Surface Normal Maps

We present a method for the registration and matching of perspective surface normal maps. Registration of two maps consists of optimally aligning their normals through a 2-D warping in the image plane in conjunction with a 3-D rotation of the normals. Once aligned, the average dot-product then serves as a match metric for automatic target recognition (ATR). We conduct an ATR experiment using synthesized views of 25 commercial vehicles, and obtain perfect recognition results when the test azimuth is within [-6deg,+10deg] of the reference pose, even when the normals are corrupted by up to 20deg uniform random noise. The results suggest that needle maps are a rich yet compact representation of an object, which may be useful for exploiting information from stereo images, shape from shading algorithms, or sensors which obtain the normals from polarization information.