BANMo: Building Animatable 3D Neural Models from Many Casual Videos

Sensitivity to inaccurate initial poses

20° pose error

50° pose error

90° pose error

We inject different levels of Gaussian noise into the initial root poses, leading to average rotation errors of {20, 50, 90}°. BANMo recovers a reasonable shape and refines the root pose even at 90° rotation error.

Sensitivity to view coverage

90° view coverage (2 videos)

180° view coverage (4 videos)

270° view coverage (6 videos)

To reconstruct a complete shape, BANMo requires all object surface to be visible from at least one frame. More videos leads to more complete shape and better better skinning weights as well as a more regular motion reconstruction.

Sensitivity to number of bones

4 bones (control points)

9 bones (control points)

16 bones (control points)

25 bones (control points)

64 bones (control points)

100 bones (control points)

We test using different number of bones for reconstruction. Using too few bones fails to recover all body parts and motion due to over-regularization. Using more than 16 bones produces good reconstructions, but consumes more memory when computing skinning weights.