CS766 Project 3 Report

Shijin Kong (krobin)

GOAL

This project aims to reconstruct 3D models from a series of images which are taken at fixed positions but with different light sources.

INSTRUCTIONS TO RUN PROGRAM

• All the work is done in MATLAB, following are four steps I have done and also described on project web page:

1. Calculate light source directions from chrome images set.
2. To reconstruct any image set, get the normal of each pixel of each image in the set, generate normal map.
3. Get albedo information for every pixel, generate albedo map.
4. Reconstruct 3D model in MATLAB

• The MATLAB source code is in file stereo.m. It does previous four steps serially and show a reconstructed 3D model at the end. To run the program, open MATLAB, change directory to where stereo.m is, and type:
  
      >> stereo(image_sets_path, chrome_set_name, object_set_name, image_count);
     
  
  image_sets_path indicates the directory containing all image sets. chrome_set_name is the chrome ball subfolder name under image_sets_path and object_set_name is the reconstructing iamge set subfolder. For example,

            >> stereo('E:\cs766\stereo\psmImages', 'chrome', 'buddha', 11);

  Two images, E:\cs766\stereo\psmImages\buddha\buddha_albedo.png and E:\cs766\stereo\psmImages\buddha\buddha_normal.png are generated as albedo map and normal map respectively.

STEREO RESULTS

• Here I show three reconstruction results as well as albedo map and normal map:

• BUDDHA (MATLAB 3D figure is here)

  albedo map	normal map
  3D model view 1	3D model view 2

• CAT(MATLAB 3D figure is here)

  albedo map	normal map
  3D model view 1	3D model view 2

• OWL(MATLAB 3D figure is here)

  albedo map	normal map
  3D model view 1	3D model view 2

DISCUSSION

It's a fun project and every image set I used is reconstructed successfully. The first three steps seem pretty fast to me in MATLAB, but the last step is time consuming. The reason is the final matrix is very large. Now we are using images with size 340x512 but it can be much larger. One possible way to solve this problem is just listing equations for those pixels marked in image mask, not for every pixel in the whole image. This will reduce the final sparse matrix greatly but at meanwhile It's more difficult to set the non-zero element in sparse matrix. Also I have to maintain the pixel position of each pixel in the least square solution. Another problem is that last step depends on continuous surface assumption, so the surface generated is smooth. If the real object has some abruptly changed corner or any other discontinuous characteristic, this method may not work totally correct.