PDF version of this entire document

Building Appearance Models in 1-D

To demonstrate the more unique results reached by a model-based NRR approach, I used the results of experiments that I described here before in order to construct models. These made use of experimental data where the correct solution was known. This data was again the depiction of bumps and the sets of images were stochastically generated with significant variability that made the problem sufficiently challenging.

Figure: 5 modes ( 2 SD) of the model built from the NRR images in Figure

Figure fig-model2 suggests that after only hundreds of iterations, the approach is able to detect - albeit to a limited extent in this case - a reasonable degree of variation, all while addressing some common difficulties. The depiction of the first 5 modes of variation of this combined model reveals that mode 1 mostly involves bump width, mode 2 mostly captures variation in height, and mode 3 captures bump position for the most part. Modes are not pure because the NRR is imperfect.

The method can handle large sets and provide a solution that does not depend on any arbitrary selection of images. Another point worth making is that a relatively small set was used here, so it is natural to expect modes to get confused and coupled with each other. I expect clearer separation given a larger set of images.