From the original inspirations of AI and vision to the aforementioned practical uses for Structure from Motion, multiple approaches have been proposed. These range from linear to non-linear, perspective to orthographic, 2-frame to multi-frame, and recursive to batch tradeoffs. The methods have their own idiosyncrasies with different input features, different accuracies, degeneracies, and flexibilities. Ultimately, the choice of which framework to use is application dependent.
We have also discussed the use of these techniques on video and continuous motion sequences and the implications that has on the algorithms. We emphasized our nonlinear recursive probabilistic approach for its stability and ease of use in these environments. These capabilities were confirmed in synthetic experiments, in real imagery experiments, under independent evaluation by others, within large vision implementations for real-time tracking, and with post-production industry evaluation. Fundamentally, this technique proved accurate, practical and resilient to a wide variety of such tests.