| Machine vision systems generally have difficulty in dealing with images of objects undergoing appearance, pose, and illumination variation. Being able to quantify such transformations can give us a potential handle to better understand and be invariant to such effects. In this work, we propose a model of such variations. Our model exploits the manifold structure generated by image data undergoing the aforementioned variations, and its connection to Lie Groups. We propose an efficient learning scheme for our model, and show experimental results on the effectiveness of our approach on both synthetic manifold data, as well as real image data from captured video sequences. Our experiments suggest that our model is robust to small amounts of noise and effective in capturing image transformations. |
