How to get better result?

[mln-wave]

Hi, Thanks for the wonderful repo, which generates 3D models for real-life photos in the wild !! I tried applying it for images of a doll. Gave 8 images as input covering different angles, in colab notebook. But results aren't good. (PFA). What can be done to improve the results?

https://user-images.githubusercontent.com/110157037/194704276-03d51202-4940-474d-99ee-7dcf7787d1a5.mp4

[jasonyzhang]

Hi,

This is a very challenging sequence. Here's what I might try to get better results:

• spend time getting the camera poses as accurate as possible
• remove the images for which the camera poses are bad. Images 3, 5, 6, and 9 have cameras that are pretty far off. This makes sense since the template shape is pretty far off from the shape of this doll, so the stage 1 camera pose optimization is probably hurting your performance. The other camera poses look reasonable.
• After stage 2 when the shape is initially optimized, update the camera poses for bad cameras manually (eg camera 5 should have its elevation adjusted). Then, re-do stage 1.
• Try a better shape initialization. Perhaps a flattened cone shape would be a better initialization that a rectangular prism. You could try using a cone mesh as the initialization, although I might also flatten it a bit in x or y dimension.

[ming-ming15]

Hi, Thanks for the wonderful repo!! How do I try using a cone mesh as initialization during testing? Are there any parameters that need to be modified?