I think if you can spare a slide on what the aim of robust statistics is, it may be helpful. The paper presents how least squares can be robustified and mentions outlier, but not the general idea.

The Jacobian and Hessian matrices in a bundle problem seems to be different and problem specific as compared to an ordinary optimization problem. It so happens that H is structured such that each entry gives a relationship between row and column indices. If, then we wanted to use third order derivatives in our system (assuming we have an optimization algorithm that uses these third order derivatives), then I think we would have to consider 3D matrices. Has anyone seen seen such an application or optimization algorithm?

Also, it is mentioned in the paper that relationships between different features (inter-feature measurements such as angles) are not included to the H in the example on page 20. But I would argue that inclusion of such relationships bring correlation problems. How would this be solved? (If anyone can point me to any reference, that would be fine too.)

A well written review. The observations on the "common misconceptions" in the vision literature and the historical overview were particularly entertaining.

I would not be covering guage freedom in my talk, maybe just give a short introduction to parametrization. My emphasis would be on sparsity structures within the bundle adjustment framework and how Levenberg-Marquardt algorithm exploits this sparsity. The content of the talk would be closer to Appendix 6 of Hartley and Zisserman (on second-order methods for iterative estimation).

The techniques and recommendations in this paper seem like they should be generally applicable to computer-vision problems, not only for calibration/pose estimates.

Will you go over section 9 on Guage Freedom for us? This was all very intriguing, but it's more information than I can absorb in the time available. I'd like to get a better feel for just how one sets up an S-transform and how to choose a good one.