3D Asset Photogrammetry - Elephant Sculpture

Elephant Sculpture

Continuing my collaborative photogrammetry projects with Logan Schwartz, we attempted to capture our first 3D asset with photogrammetry. We picked this cool elephant sculpture because we liked the surface qualities of the wood and the carved details. Our goal with this mini-project was to create a 3D model that I could texture by hand to try to replicate the original.

We gathered supplies such as my Canon EOS 7D, a circular polarizer, a small tripod, backdrops, and a color corrector. The weather was overcast and snowy so we had near-perfect lighting conditions. The wood is very reflective so we knew that we had a challenge ahead of us. Using the circular polarizer, we tried to cancel out as much of the bright reflections as we could, as well as some assistance from editing the photos in post using Lightroom.

Instead of rotating the camera we rotated the object (which causes issues later in Reality Capture) in order to avoid capturing unnecessary background data. We were skeptical if it would work or not but we thought it was worth a try. We captured a medium height (straight on) rotation, as well as a raised height rotation to try to get a different perspective of the object.

Results

Putting the photos into Reality Capture was surprisingly easy, with step by step instructions built into the program in case you get lost. We were also surprised by the level of detail captured in the scan data despite some imprecise level of error in our photo taking process.

Initial Reality Capture Fail

When the photos were first brough into Reality Capture, the mesh did not generate correctly. It took several attempts but eventually the program started to understand the photos better.

Improved Reality Capture Mesh

The generated mesh was not perfect and had a lot of issues such as overlapping 3d model data from photos facing opposite directions. Artifacts such as multiple overlapping legs and copies of the mesh around the viewport plagued the file. These issues stemmed from rotating the model for photos rather than rotating the camera, which created minor misalignments each rotation increment. This created a model output that was impressively detailed on the front, however devolved into a low-poly geo mess on the backside. In order to resolve this issue, we determined that when we attempt to recapture this object, we would try both capturing photos around the perimeter of the object, as well as creating a turntable the object can sit on to rotate precisely centered. I'm not sure which method will produce a better result but I look forward to trying it.

However, I was very impressed by the level of detail on the model. I had originally thought that the program would not pick up on fine surface details such as the carved lines on the legs or the face, it replicated them with precise accuracy.

Another issue we ran into was the backdrop color bleeding on to the model. Very little color information about the wood actually showed up in the textured file besides a general reddish-orange hue, contaminated by the blue backdrop. In order to resolve this issue in future attempts we have decided to try again with a white sheet backdrop in order to minimize color spread on to the asset.

Arnold Render

I brought the model into Arnold to render a still image of the model. While it is mostly a failure, we learned a lot from this attempt. Capturing 3D assets proved to be a lot more difficult than capturing flat textures. I was very excited to see the 3D version of the geometry in Maya though and I have a strong desire to try this object again as well as other objects.