3D PHOTOGRAPHY

The Head

by

Dragos A. Harabor

Main Steps:

• Calibration
• Coordinates Recovery
• Triangulation

This Year's Approach:

• Alignment
• Volume Carving and Marching Cubes

Last Year's Approach:

• Mesh Creation
• Unshading Mesh (?)
• Mesh Zippering

CALIBRATION:

• used checker board pattern on a flat plane
• things to try in the future:
• use a pattern on two planes that form a well known angle (90 deg. for example)
• why? might help get better results, make sure that we do not run into problems like the plane is not a plane, the angles are not preserved, etc.
• problems:
• quality of the equipment:
• the projected pattern/stripes flicker
• higher resolution camera means more details
• precision in determining the corners of the squares
• least squares might end up in local minimum
• overall rating: GOOD

COORDINATES RECOVERY:

• in the camera frame of refrence:
• can look only at the boundaries from the finest stripes image or use all the boundaries => more points
• tried both; conclusion:
• get more points, which is good
• get more noise, which is BAD; solution:
• have an editor to eliminate the noise
• intermediate solution (i.e., get points only from let's say half of the images)? maybe.
• in the projector frame of reference:
• use GREY CODE
• add 0.5 => big source of confusion; do NOT do it in the future, or at least be consistent; yes, I understand why 0.5 is added!
• a lot of noise:
• why? mislabeled points.
• solution: point editor.
• estimated number of points (for my statue): 33,000 pts. per view; that's a lot in terms of processing time!
• overall rating: VERY GOOD, provided good data to start with.
• images

TRIANGULATION:

• just pure math; make sure you get the right formulas
• do NOT use for loops in Matlab and it will be fast even at 30,000 pts.
• noise again; solution:
• you are actually right: noise editor.
• images

ALIGNMENT:

• tricky!
• 3dpalign still needs work, especially when it comes to sliding views
• do not use too many views, but only the necessary ones => it will pay off at the next step
• be prepared to spend a lot of time => go for seconds (second set of data I mean...)
• overall rating: OK
• images

VOLUME CARVING and MARCHING CUBES:

• the most interesting part of the project; why?
• find about all the previous failures and mistakes:
• calibration
• alignment
• right resolution?
• probably 1024x1024x1024 => get a lot more details
• 256x256x256 requires 32Mb disk space and some procesing time
• going up to 1024x1024x1024 multiplies the above with a factor of 64
• play with the parameters: smoothing factor, isosurface value, gaussian distribution
• software rating: EXCELLENT
• images; do not look very well because of the tiny errors in calibration and alignment.

MESH:

• fill in the holes, by looking at the neighbors: 4 valid points seems to be the right number (for me!)
• smoothing is very important: for all the points average the value of the point and its valid neighbors
• generate triangles using the shortest diagonal
• apply texture
• get a big VRML file
• uses DECAL model which is not very realistic
• images

UNSHADING THE MESH:

• advantages:
• decrease the size of the file by 20%
• use MODULATE model, which is more realistic
• use a new orthogonal texture map that needs to be generated in advance, accounting for Lambert's cosine law
• it really works!
• images

GENERAL COMMENTS ABOUT THE METHOD:

• it seems to work if the data to start with is good enough
• what's "good enough"?
• accurate data => high quality projector and camera
• careful points recovery
• very good alignment
• marching cubes at high resolution (1024)
• processing time: 1 day if you are lucky!
• disk space: a lot is still not enough
• IDEAS:
• try it with a color camera
• maybe we should use a different calibration method (see above)
• alignment:
• make it more sensitive to sliding
• user interface:
• should be able to align two or more views and then lock them
• parameter for the number of times to run the global alignment step
• should be able to translate a view, not only rotate it

ABOUT THE CLASS:

• enjoyable class because:
• EXCELLENT class
• extra topics, beyond the necessary information for the lab work