Optimization:
    Support for IPP (ippcv, ippi, ippvm ...), MKL/atlas

SFM and 3D:
    Higher-level Optical Flow:
        (verification via epipolar geometry reconstruction
         and backprojection error measurement - ransac etc.)
    Tri-focal tensor

Comp. Geometry:
    - Min-eps contour approximation (fixed number of sites)
    - Graphs from point sets (NNG, MST, RNG ...)
    - CvSubdiv2D -> CvGraph conversion
    - More contour comparison functions (e.g. Housedorph distance)
    - Primitive geometric operations (test pt-contour,
              contour-contour intersection, union etc.)

Misc. Functionality:
    - Smarter image segmentation methods (?)
    - Textures handling to main library (better specs, methods?)

Documentation:
    Document HaarFaceDetect

Apps:
    Rewrite demos in FLTK(?) to make them cross-platform

Interfaces:
    C++ interface (STL-like interface to CvMat
       (see cvaux/include/cvmat.hpp as a starting point), CvSeq,CvGraph)
    Python, Java etc. interfaces using SWIG

Cameras:
    Make a single camera module instead of 3. That is what we have now:

       Module      Interface                Win32                    Linux

       cvcam       C (callbacks)      DirectShow (AVI/Cam)         V4L[2] (Cam)
       highgui     C (no callbacks)   VFW (AVI/Cam), MIL (Cam)     ffmpeg (AVI)
       camera.cpp     C++                  VFW (Cam)                   -

    The proposed variant:
       highgui like interface (i.e. no callbacks),
       extended with some cvcam features like customization dialogs,
       resolution changes, camera selection etc.),

       Win32: DirectShow, VFW, MIL
       Linux: FFMPEG + V4L2

Others:
    Split cv for better modularization (core, image, geometry, 3d, motion)
    Throw out or improve old stuff (dense opt. flow, 2d hmm,
                                   eigen objects (simplify interface))

    Better test system with regression tests (improve trs(?))
    Performance (speed) tests