Online globally consistent mosaicing using an efficient representation

This work addresses the problem of mosaic construction from image sequences. We propose an incremental, globally consistent method in which each new frame of the sequence is aligned to the mosaic that has been constructed up to that time instant. We propose an efficient, stochastic mosaic representation, in which the color of each pixel is represented with a mixture of Gaussians. The parameters of the Gaussian mixture are learned on-the-fly. The proposed method is particularly suited for long image sequences, since it prevents accumulation of small alignment errors and maintains constant memory requirements. Results for synthetic and real image sequences are presented.