February 08, 2014

Take a more top-down approach; split into "Modeling" and "Inference" sections. give details on request.

Tailor more toward committee audience and less to a general audience.

Questions to ask yourself:

- What is the purpose of this talk to this audience?
- What do you want the audience to remember?

Structure the talk and paper more tree, with details on the leafs.

Get the the Arabidopsis application faster.

Have a "map" of the various research parts; show how they connect and where we are currently. **Need a stronger big picture.**

Focus on the key pieces of research, ditch the rest (e.g. drop DTW focus).

Related work (SfM and SfS). I was worried I left out too many recent developments, but no one commented on it (but it's in the paper).

Need stronger transitions between topics.

Unclear how much left there is to do.

Unclear what the key contributions are, and what is just "other stuff we did".

Too much time talking about alternative approaches. (I can probably cut quite a bit, time-wise by taking this advice).

Graphics: backprojection lines too light.

Phenotyping

projective geometry (a few good pictures would be better)

Linear approximation details (e.g. emphasis on running time).

Temporal component of model is not for camera motion, but for trackingnonrigid deformation.

3-tuple slide: add a ground plane to images to convey 3D, maybe eliminate hard black outlines what falsely convey 2D image-plane context.

How is DTW section connected to the rest?

When I say parts are "given" or "assumed", what does that mean? i.e. how are they given? (This should be clearer when I re-orgainze top-down. Natually handled in inference)

What does it mean that CS model is more "expressive" than squared exponential model?

Likelihood linearization was confusing, esp. the justification. Since local linearization isn't novel, could just say we linearize, and exploit conditional independence of points to do it in linear running time.

In paper, mention that index estimation is currently slow and ucnlear if fit for MCMCDA approach.

What constraints are there on data collection? Are cameras required to be on a turntable?

What is the connection between the covariance function and the species? Do you think the cubic spline covariance function this is generally applicable?

What is the connection to the science? What can we learn about species from this? (I could probably talk about two aspects - (a) can build models using data recovered from 3D reconstruction and (b) learning population models in heirarchical BGP (phenotypes as parameters))

Posted by Kyle Simek