CVPR 2014 Summary
Summary of completed work, work to be done.
Summary
- Developed analytical methods for marginalization/maximization that working with rank-deficient precision matrices.
- Improved inference of index-set
- Trained model
- Attachment functionality
- Developed per-view model (temporal model)
- Re-formulated attachment in terms of GP kernel, instead of a matrix operation* Incorporated non-gaussian pixel model (opengl silhouette rendering, CUDA)
- revived, refactored old silhouette rendering code, blurred-difference edge likelihood code
- Fixed errors in ground-truth, finished incomplete ground-truth
- Ran reconstruction on all ground-truth
TODO
- Address problems with ground-truth reconstruction
- Improve 2nd likelihood speed.
- End-to-end sampling
- Add new likelihood
- Split/merge
- Idea: bias using foreground/background pixel probabilities
- Evaluation
- Find 2nd Dataset (neurons?)
- Add boost_system to KJB build system
- Handle non-constant curve widths