A Matching Shader Technique for Model-Based Tracking

We present a line feature matching method for model-based camera pose tracking. It uses the GPU for computing the best corresponding image line match to the edges of a given 3D model on a pixel basis. Further, knowledge about the model is considered to improve the matching process and to define quality criteria for match selection. Each edge is rendered several times with image offsets from the last estimated position of the model. The shader counts the number of pixels in an underlying canny-filtered camera input image. Returning the best fit by pixel count can be done applying occlusion queries. A speed-up can be achieved using a more elaborate shader with texture read-back reducing the number of rendering passes. The matching shader is not limited to work with lines and can be extended to other structures as well.