Progress Report
February 28th, 2006
Overview
- Conference decision, preparation, submission plans
- Papers of interest (cyclic journal browsing)
- Monday Talk
- Revised Form 12 (including appendix)
- Entropy-based NRR assessment
- 3-D Extensions
- Correspondence and other related matters
Paper of Interest #1
- The following 2 papers seemed somewhat relevant to our work
- The first one is due to be published in TPAMI, April 2006
- Robust Point Matching for Nonrigid Shapes by Preserving Local Neighborhood Structures
- Abstract:
- "In previous work on point matching, a set of points is often treated as an instance of a joint distribution to exploit global relationships in the point set. For nonrigid shapes, however, the local relationship among neighboring points is stronger and more stable than the global one..."
Paper of Interest #2
- Validation of Bone Segmentation and Improved 3-D Registration Using Contour Coherency in CT Data
- To be published in IEEE TMI March 2006
- Abstract:
- "A method is presented to validate the segmentation of computed tomography (CT) image sequences, and improve the accuracy and efficiency of the subsequent registration of the three-dimensional surfaces that are reconstructed from the segmented slices..."
- IEEE journals now delivered as RSS feeds, second to do so after ICCV
Form 12
- Revised progress summary/outline in Form 12
- Thesis struture/plan changed in accordance with last meeting
- Revised version sent be E-mail
- Meeting in middle of March, around the same time as the MICCAI deadline
Monday Talk
- Delivery even worse than the last one
- Needed more rehearsal (for better familiarly with slides)
- Slight distraction due to work on 3-D extensions/port
Conferences
- Discussion of possibilities throughtout the previous meeting
- MICCAI - possibly description of the entropy, a draft exists
- MIUA - content can be made similar to that of the ISBI paper
- BMVC - a CVPR 'fallback'
- ISBI - need to book flights for the group, verify attendance, funding sources
Work on Entropy
H(G[S->T])
is problematic: synthesis to training set
H(G[S->S0])
is more acceptable: synthetic set to its subsets
- Need normalisation that involves
H(G[T->T])
- No substantial progress in that area, primarily due to 3-D extensions taking precedence
3-D Extension of the Methods
- For 3-D NRR assessment, need data volumes (we have about 200 of them)
- Vlad or Tim could build models from subset (making it more feasible)
- A large registration process is involved, but is a 'one-time occurrence'
- 3-D registration gives triangulation, used to produce a model
- Synthesis from a 3-D appearance model, then can say something about Generalisation, Specificity and Entropy (more normalised, principled)
- Experiments can take plenty of time to test, maybe weeks or months
3-D Extension: The Algorithms
- Analyze format input
- Shuffle distance or Euclidean where shuffle uses a neighbouring cube
- Estimates: a few minutes for a single Euclidean comparison, under an hour per image comparison with large shuffle radius on Pentium 4 workstations
- Proposed set size: 10 or 20 images in training set, 100-1000 syntheses
- 3-D model evaluation can be parallelised and run in clusters overnightas
- Proof of concept rather than proof of practicality
Miscellany
- Prolonged contact with academic in France
- ASM, AAM, MDL and their use in 3-D for segmentation
- Interested in the work being done in Manchester
- Postgraduate student in other department needs help with edge detection