A Probabilistic Notion of Camera Geometry: Calibrated vs. Uncalibrated

We suggest altering the fundamental strategy in Fundamental or Essential Matrix estimation. The traditional approach first estimates correspondences, and then estimates the camera geometry on the basis of those correspondences. Though the second half of this approach is very well developed, such algorithms often fail in practice at the correspondence step. Here, we suggest altering the strategy. First, estimate probability distributions of correspondence, and then estimate camera geometry directly from these distributions. This strategy has the effect of making the correspondence step far easier, and the camera geometry step somewhat harder. The success of our approach hinges on if this trade-off is wise. We will present an algorithm based on this strategy. Fairly extensive experiments suggest that this trade-off might be profitable.