I am using photogrammetry to capture a series of objects for scientific analysis, so my goal is to create a reliable method and be able to explain any anomalies.
I have a very consistent set up, with studio lighting, a decent SLR and lense, and a turntable. Objects (5-20cm across) are placed on the turntable and rotated with the camera on a tripod. I take between 120 and 300 images of each object depending on size and shape. Normally the reconstruction is good, but yesterday two reconstructions failed with either a large fuzzy sparse cloud, or a long line of points stretching off into infinity and all of the cameras clustered around a single point. (I am using masking).
After playing around with these for a few hours, I could not get them to align correctly. I tried different alignment options, different KPL values and accuracy settings - nothing helped. I reset all cameras and tried realigning them but also no good - this did stop the "shooting off" into space, but it kept resulting in a fuzzy point cloud, or with a second "ghost" point cloud of my object slightly offset from the other
I looked to see what the difference was between these scans and the 10 others that did work, and realised that in this scan, I had flipped the object once during the middle of the scan as I was unable to move the tripod at the time. My usual method is to take rings of photos moving the camera down at each ring, then flip the object and take more rings moving the camera up each time. These two I took 1 ring, flipped the object, took another ring, then moved the camera down and repeated.
With this in mind, I reset alignment on all cameras and then manually did them in the order as if I had taken them using the normal method. The alignment was perfect.
I repeated this several times on both scans and can confirm that this results in a near perfect reconstruction, where just processing them in order results in various bad reconstructions.
The order of the photos is obviously very important to the algorithm used for reconstruction. My questions are:
- What is the optimum order for the photos to appear in to help the matching?
- Does image rotation matter (I had a previous question on this), ie if I rotate the images with the object "flipped" does this help the reconstruction?
- What is Photoscan using to order the photos for processing: filename, date and time, order of photos in the grid within the application?
- Is there anything else I can do to "help" the reconstruction algorithm?
Sorry for the long winded post, thanks for your help.