代做CISC 4/642、代写Matlab程序、代写C/C++、Python设计

日期：2018-11-22 10:16

UNIVERSITY OF DELAWARE

DEPARTMENT OF COMPUTER AND INFORMATION SCIENCES

CISC 4/642: COMPUTER VISION

Programming Assignment-2

For 642: Due 11/29/18 (10 pm)

For 442: Due 12/7/18 (10 pm)

Design and implement Matlab or C/C++ or Python program for a stereo analysis system

involving feature-based, region-based and multiresolution matching. The program should be

able to perform multi-resolution stereo analysis, where the no. of levels are set by the user.

The template and search neighborhood at each level can be set differently by the user, so is the

method and matching measure to be used.

Details below.

1. Methods

-- Region-based analysis

i) Give the user options to pick template size and matching window. Template can be a

rectangle; matching is 1D looking in only one direction for the match.

ii) Provide the following matching scores: i) Sum of Absolute Differences (SAD),

ii) Sum of Squared Differences (SSD), and iii) Normalized Cross-correlation (NCC).

_ Feature-based analysis

Use Harris corners extraction for feature detection, then use the descriptor value for matching

(i.e, Harris corner response measure). Also give the user option to choose different matching

scores: SAD, SSD, NCC.

2. Stereo Matching

i. At a given level, stereo matching is done and disparity calculated.

ii. Validity check is then performed, wherein if the left-to-right match does not correspond to

right-to-left match, a 'zero' is placed at that location in the dispariy. One way to do this is, switch

the left image with right image and see if the correspondences match up in both directions. They

won't match up, for example, when pixels are under occlusion (see Fua's paper provided).

If the correspondences are same in both directions, then consider that the correspondence are

valid, otherwise invalid.

iii. Perform averaging in the neighborhood to fill these gaps (zeroes). The neighborhood

window can be as big as needed (for example, if 3X3 does not provide atleast 5 non-zero values,

make the window 5X5, and so on). Some may use quadratic fits in this neighborhood to get the

interpolated values in the gaps – I will leave this upto you.

iv. Propogate disparity to the next lower (finer) level. You may duplicate disparity from 1-pixel

to the corresponding 4-pixels in the lower level, or only transfer 1-pixel into the lower level, and

then perform interpolation to get the missing (three) pixels at each location.

v.Use this disparity as starting point of the search and perform stereo matching and update

disparity.

3. Data

Please visit below page for stereo data and evaluation strategies:

_ http://vision.middlebury.edu/stereo/data/

4. Notes

_ you don't need to create left and right images at each lower level from the higher level, instead

create from the full resolution images (or store it somewhere while creating higher levels).

_ you may end up with sub-pixel values of disparity while interpolating. This is just fine, and

good.

_ Better your stereo analysis results, better will be your grade. So you may incorporate your own

constraints and improve in anyway the accuracy of the algorithm.

_ Prepare atleast 5 results, with a word file explaining each result and indicating parameters

used; upload the 'tar' directory to Canvas which includes source code, data used. Include all the

necessary files.

_ You may use corner detectors and image warping code available on the web.