Hello,
I am currently evaluating methods for 3d-documentation of archeological objects.
for this purpose i built an arduino controlled turntable as well as a sketch for timed shutter releases, and i would like to share my results with the community.
please feel free to point out any blatant mistakes on my side
the setup:
-simple cullman tripod with two flexible clamp-on desk lamps (equal bulbs)
-arduino uno
-adafruit stepper shield
-relay module (for the shutter)
-canon eos 1100d, kit lens
-hama cable shutter control
-lenovo thinkpad edge A18
things that i found to be adamant for successful models:
1. even lighting of the object in the camera's fov. this includes the barely visible sides of the object!
2. the masking can bit a little painful at times, so i chose a backgrond color that contrasts with the object. if you use textiles, make sure the cloth is even and flat (the shadows make masking a real pain)
2. if you mask in Photoshop:
- create a layer with low contrast and high brightness (makes selection easier, can be deleted before masking)
- select the background instead of the object (easier and more accurate than masking an object with heterogenous surface)
- set all settings in the "refine edge" menu to zero. this gets rid of floating artifacts around the finished geometry
- use prerecorded actions so you can automate the masking. all you have to do manually is the selecting
3. number of fotos
i initially took 32 fotos per turn per camera-angle using 4 angles, this had me at 128 fotos total.
after allignment i was able to delete almost half of the fotos without losing much of the points, leaving me with 73 fotos. this resulted in greatly improved proccessing speed.
4. set your apperture to a high value (i.e. closing it) to enhance the depth of focus, crank up your ISO and lower the shutter speed. yes, you get more sensor noise that way, but photoscan seems to cope better with that than with parts of the object beeing out of focus.
By
enkidu83 at 2012-06-16
By
enkidu83 at 2012-06-16
5. if you are doing flat objects (like reliefs or clay tablets) where the essential spatial information you are trying to collect is on a single surface you should try following technique:
- project a random grayscale noise image on your object, focus your projector
(don`t touch the projector from this point on!)
- adjust your camera's iso and apperture so that your pics look like the photo in pic 3 in the attachment
- keep the shutter speed high so you can shoot without tripod
- take about for pictures from different angles (freehanded), always capturing the projected surface as a hole (this way photoscan doesn't have to look for pairs)
- mount the cam on a tripod
- take one last picture with projected noise, use a remote shutter release (picture "A")
it is very important that you don't move the camera or else this won't work!
- shut off the projector
- take normally lit picture from the exact same angle and position (picture "B")
- feed all noise-pictures into photoscan (including "A")
- set "maximum points per foto" to 40000
- generate pointcloud and geometry (should only take an instant)
- save project, close photoscan
- in your folder, replace picture "A" with "B", (don't forget to rename B to A)
- open Photoscan
- generate texture from single foto, choose foto "A" (i.e. "B")
this method is incredilby fast and the results are stunning.
last pic is shaded in meshlab using the zbrush-shader. no postprocessing.
when using a high shutter speed the color of the projector will be off (due to the projectors framerate), but since photoscan grayscales the images for point-cloud generation it doesn't seem to affect the results.
By
enkidu83 at 2012-06-16
By
enkidu83 at 2012-06-16
By
enkidu83 at 2012-06-16
any input and criticism is welcome
Show Posts
