More Work Being Done

I've taken a few more photos to create photosynths. This time, I focused on having even lighting all over the objects I was taking photos of. In total, I had five light sources - two about fifty cm away either side of the object, and three ceiling lights within 2m of each other. This helped to eliminate the issue I previously had of my shadow being cast on the object when taking photos from certain angles. The shadows would cause the object to take on a slightly different colour in the photograph and reduce the probability that an accurate point cloud would be formed, so it definitely helped to reduce the impact of that aspect.

For the first two sets of photos I took (which were of the Gundam model), I had the object sitting on top of one of the light sources to see if being directly adjacent to the light would help increase the photo set's synthy percentage. It resulted in the sets being 100% synthy, but the point cloud turned out to be mostly on the brightly lit paper below the model, rather than the model itself. Here are the results:





After these tests failed, I decided to move the light from below the object so it would be sitting about fifty cm away from it. This worked much better for creating accurate point clouds, but the synthy percentage was lower than before. Given the point clouds were more accurate, I decided to keep this lighting layout for the remaining photo sets.

In addition to this, I wanted to try getting an accurate point cloud of the Piccolo figurine I had tried several times before. A few issues with the figurine in previous attempts were that it has diffuse reflections on most of its surfaces, and that its surfaces also have homogeneous rather than heterogeneous textures. In a bid to address both of these issues, I covered it in talcum powder and draw across its folds and wrinkles with a dark pencil to make them even darker and more distinct from the rest of the adjacent surface.

This did improve the figurine's synth percentage substantially (51% and 82% for each photo set), but the resultant point clouds were still not detailled enough to produce an accurate model in Meshlab (I'll edit this post and add images of the resultant model when I can get on my laptop to screenshot them).

Here are the two synths I described above:





Finally, I decided to try getting a better Photosynth of my NES controller. An issue with its previous synth was that it has many symmetries due to being a rectangular prism. The back of it synthed very well because it was heterogeneously textured with scratches, but the sides, top, and front didn't synth well at all because they're very homogeneous in shape and texture. To help resolve this issue, I tore up some paper, wrote large letters on the pieces, and stuck them on the problem sides. As an afterthought, I also scribbled on the controller with pencil for good measure.

The first set of photos for the synth was made up of a 360 degree spin around the controller at a constant height, and rotating up and over the controller (starting with a side view, then finishing with a top view).

The second set of photos included the above set, and also two sets where I kept the camera pointing in one direction at an angle to the front of the controller, and panned it past to try and capture detail to do with the controller's buttons and D-pad. I tried this because I wanted to create a point cloud that accurately modelled the buttons, rather than turning them into somewhat amorphous blobs.

The results of the photosynths were much better than before, producing very detailled point clouds. However, unfortunately, the buttons didn't show up very clearly in Meshlab even with "Octree Depth" raised to 12 in the "Surface Reconstruction: Poisson" dialogue box. Here are the results of the photosynths:

From the first set of photos:


From the second set of photos:


The point cloud I ended up using was the one from the first set of photos, since its point cloud appeared to be more accurate. Here is the model as a point cloud in Meshlab, and as a mesh after it was exported to 3DSMax as a .obj file:







Note in the lower-left image that the A and B buttons appear to stand out on the mesh, as well as the location of the indentation and bumps for the Start and Select buttons. The D-pad, however, seems to have resolved into a slightly diagonal bulge. Though the point cloud looks accurate when textured with the colours of the feature points, it seems that the crisp edges of the buttons and D-pad haven't been captured. Or, perhaps, the method Meshlab used to reconstruct the points resulted in smoothing out what it perceived to be imperfections in the point cloud, when in fact the points around the buttons should have had sharper angles between adjacent polygon planes.