# Title and Due Date
Lazy Snapping: Optimizations and Extensions – Due February 23, 2006

# Problem Statement
The current implementation of Lazy Snapping has severe bottlenecks which
prevent it from being a truly real-time and interactive method for
performing image cutout.  Several parts of the system need to be optimized
or implemented differently.  Lazy Snapping also lacks some features which
would improve usability, specifically editable cutout borders.  Finally,
Lazy Snapping has room for improvement in terms of speed, accuracy, and
interface.  We need to fine tune variables such as minimum segment sizes
and edge weights in order to assure the best results and may be able to
eliminate the need for user  interaction by incorporating saliency into
the algorithm.

# Methods
The segmentation of the image can be improved by replacing the mean shift
algorithm with the Watershed algorithm.

The accuracy of Lazy Snapping can be improved by allowing users to edit
the cutout border before performing foreground and background separation. 
This can be implemented with a similar method as that described in the
second half of the Lazy Snapping paper.

It may be possible to speed up the performance of Lazy Snapping by pushing
some work onto the GPU.  The plausibility and necessity of this are
unexplored, but can be kept in mind as the project continues.

Saliency may be a suitable replacement for user identification of
foreground and background areas of an image.  Segments with high saliency
would be automatically considered foreground and those with low saliency
would be considered background.  We are unsure as to what the results of
such a system would be, but feel that saliency should allow us to either
remove user interaction all together or at least prove a useful addition
to  the algorithm when used in conjunction with image marking.

# Plan
Milestone 1: Replace mean-shift image segmentation with the Watershed
algorithm.
Milestone 2: Optimize Lazy Snapping by improving clustering and graph cut.
Milestone 3: Add Lazy Snapping to the existing image library.  This may
include some restructuring or improvement of the image library.
Milestone 4: Improve the Lazy Snapping UI by implementing an editable
cutout border which allows users to fine tune the results.
Milestone 5: Research the effects of using saliency for seed values.

# Evaluation Criterion
Lazy Snapping should be able to operate at interactive speeds.  Users
should be able to load an image, easily separate foreground and
background, quickly receive feedback on the results of the algorithm, and
edit those results until they obtain a satisfactory and accurate result. 
Our results should be similar in quality to those present in the Lazy
Snapping paper.

Lazy Snapping should be integrated into the image library in a modular and
easy to use fashion.  Each part of the algorithm should be a separate
implementation which could be used on any image for any purpose. (i.e.
Separate image segmentation, clustering, graph cut, etc.)

The results of our research questions, including saliency and the
difference between max-flow and shortest path graph cuts, should be
answered with measurable counts of speed and quality.

# Deliverables
A stand alone Lazy Snapping application which allows for easy image
cutout.  This should give the user control over how the image cutout is
performed and allow them to edit results to their liking.

The implementations of image segmentation, color clustering, graph cut,
and lazy snapping with good documentation and easy to re-use code.  These
will be combined with the image library for future use.

# Readings
I have already read...
1)Lazy Snapping
2)Interactive Graph Cuts for Optimal Boundary and Region
3)Watershed Segmentation in Digital Image Processing
I will read...
1)More material on Watershed image segmentation
2)Papers on image saliency
3)The current Lazy Snapping code
4)The ITImage library
5)Looking for suggestions...

# Risks
Many of our goals have arbitrary time constraints as they focus on
optimization and speedup.  This could lead to work taking longer than
projected or needed and result in pushing many goals to the end of the
project time line.  If we want to avoid a crunch at the end of 5 weeks, we
will need to stay focused on our long term goals and ensure we don't get
caught up in early optimizations.

# Motivation
Image cutout is a complex problem which has been solved with some success,
but still contains much room for improvement.  A fast and accurate method
to perform image cutout would be useful not only for image editing
applications, but also in other image processing work.  If saliency proves
a useful method for performing image cutout without user interaction, many
other applications could benefit.