January 06, 2014

Spent morning and early afternoon doing preparation for my dissertation proposal by doing some background reading and laying out a rough sketch of what I'd like to cover.

Read Ernesto's dissertation proposal. The bulk of Ernesto's proposal seemed to be his two published papers, with an extended literature review and a brief description of remaining work to be done. This isn't too surprising, since he was so far along in his research when he wrote the proposal. Unfortunately, I have no such publications to draw on, but I may be able to use part of the tracking paper, which I was third author on. But on the whole, I'll need to lean more in the "proposal" direction, and less in the "dissertation" direction than Ernesto was able to. Time to dig for more representative examples of the proposal I'll be writing...

Read this article, describing a dissertation proposal in science. The second page has a nice organization of possible sections for my proposal.

According to the aforementioned article, length is 10-40 pages. This is supported by Ernesto's proposal (25 pages w/ references), and these example proposals from computer science, which fall between 9 and 20 pages, with references. The number of references is between 14 and 85, with a median around 20.

I was surprised to see the level of brevity and abstractness in most of the example proposals above. I'm assuming these were written early in the research phase, shortly after the end of coursework; it is encouraging to see that this document need not be a *tour de force*. Overall, I'm thinking that since I've developed my research so extensively, my proposal will probably be heavier on references and detail, and likely longer than the average. However, I need to avoid falling into the trap of trying to write my dissertation instead of a proposal.

I'm feeling more confident now that I can finish this by the deadline I set for myself of January 31, and a reschedule hopefully won't be needed. My initial investigation suggested that writing the proposal should take three to nine months, a surprise that knocked the wind out of me. I realize now that much of that time is spent investigating topics, doing initial research, and reviewing literature, not specifically writing. Since I've completed those steps (extensively!) it should be reasonable to assume I can write the full document in 2.5 weeks. I'd like to get Kobus a first draft within the week, with at least the rough structure layed out, so he can correct my course if I'm way off.

I briefly reviewed the computer science department's graduate program policy for comp exams (something I haven't looked it in probably far too long!) and realized I should have scheduled my dissertation proposal in Fall 2011!! It should be no surprise the department was urging me to complete this right away!

Below are notes I jotted down while reading Ernesto's proposal.

Research items worth covering in the proposal (or dedicating chapters to in the dissertation proper)

- Edge extraction, Stroke-width transform skeleton.
- Bayesian Model
- GPU-enabled edge-based likelihood
- Blurred-difference likelihood
- Chamfer likelihood
- GMM likelihood

- Prior: "Branching Gaussian Process"
- Modification for multimodal likelihoods (Importance sampling)
- novel covariance function
- Temporal modeling

- Marginalization: Laplace Approximation

- GPU-enabled edge-based likelihood
- Initial estimate: Dynamic programming algorithm for multi-view triangulation
- Inference: MCMCDA
- Index estimation
- Analytical gradients derived
- Dimensionality reduction? (for pixel likelihood)

- Multi-view Reconstruction
- Visual Hull
- Voxel-based (Hough transform)
- Space carving (Hough + photoconsistency)

- Biological application
- Root structure Architecture papers
- Neuron Tracing
- Vascular segmentation and modeling
- Tree branches - Amy Tabb, CVPR-2013 [pdf]
- L-systems - Metropolis Procedural Modeling
- S-C Zhu plant paper - Bayesian reconstruction of 3d shapes and scenes from a single image

- Tracking
- Oh et al.
- Ernesto
- Tracklets? (TODO)

- Bayesian Model Choice
- P.J. Green, RJMCMC
- Laplace approximation
- Candidates estimator
- Examples without model choice?

- Curve models
- Song-Chun Zhu - Parsing Images into Regions, Curves, and Curve Groups
- snakes
- Curve indicator random field
- Implicit definitions

- Edge-based likelihoods / energy functions
- Curve indicator random field
- Joe Schlecht's papers
- Chamfer matching
- Shotton, PAMI 2007 - Multi-Scale Categorical Object Recognition Using Contour Fragments
- Shotton, ICCV 2005 - Contour-Based Learning for Object Detection

- Poon & Fleet 2002 - Hybrid Monte Carlo Filtering: Edge-Based People Tracking
- TODO: More examples of edges in bayesian inference
- Gradient Vector Flow
- Blurred Gaussian

- Thin plate splines
- Ferarri - Accurate Object Detection with Deformable Shape Models Learnt from Images
- TODO: others

- Deformable models
- Fua, Uratsen (TODO)
- Monocular Template-based Reconstruction of Inextensible Surfaces Perriollat, Hartley, and Bartoli

- Sampling-based Structure From Motion
- Dellaert - SfM Without Correspondence (and other?)
- Forsyth - Bayesian Structure From Motion; Joy of Sampling
- TODO: dig here

- Misc Structure from Motion
- Semantic Structure From Motion?

- Model-based Reconstruction
- Joe, Luca's work
- Savarese, Fei-Fei - 3D generic object categorization, localization and pose estimation
- TODO: dig here

- Evaluation
- Diadem Challenge

- Gaussian Process
- Basic: Williams and Rasmussen
- GP-LVM
- Urstein Ulenbeck process

- Miscellaneous
- Stroke width Transform

An initial strategy for laying out the proposal is below. I may want to rethink this after reading this guide, which suggests a much higher-level strategy. However, the level of detail in the strategy that follows may be justified by the late stage I'm at with my research; could contribute to a stronger argument. Should sleep on it.

*Intro.* describe problem, motivate and give background.
*Body.* One Seciton per part, with one or more of the following parts

- Background and related work (show it's sensible and proven, but also novel in this case)
- Progress so far (include derived equations, intermediate results, etc.)
- Work to be done (defend why it's promising).

- Leafs, flowers?
- Root structure architecture?
- Neurons?
- Vascular modeling
- Likelihood improvements - color, patch-based?

- Camera setup
- One camera, multiple angles
- Multiple cameras, many angles

- Degenerate likelihood
- (rough diagram in notes)

- Data / edges
- Likelihoods
- BD Likelihood
- GMM per-pixel

- BD Likelihood

- Two-phase likelhood
- Curve Covariance - 1. smooth, 2. Rate and offset, 3. temporal
- Branching covariance
- Laplace approximation
- Marginal likelihood approx w/ Laplace
- Mean approx w/ Laplace
- Index gradients

- Think more on diagrams; what do I have? what do I need?
- Width as depth-cue; single-view reconstruction
- GPU-based gradient? Can render index w/ each edge. tell GPU how index moves to avoid re-rendering full model.
- Is mean-offset really a problem? If there's no size/shape distortion (pure translation), maybe not in many applications.
**Use chicken/egg for index optimization in the wild**. As opposed to energy minimization we use for ground-truth. The data will drive the fitting well enough so clever index fitting isn't necessary.- Remember to draw from the blog (TODO)
- Add MRF smoothness to pixel Likelihood? (to address non-dependence)

- Contact Dr. Zhang or Dr. Gniady about joining the committee.
- Complete form in UAccess.
- Look for quality blog posts to draw content/equations
- Lit Review
- Look into tracklets; MCMCDA? Ernesto has a referencek
- examples ignoring model choice problem?
- More edge-based likelihood applications
- Ferrari papers?

- Dig into (Vladlen Koltun's literature)[http://vladlen.info/] on reconstruction.
- Dig into Ferrari edge/contour literature.