Oriented bounding boxes are a useful bounding primitive, used for example in Newton.
Newton currently computes them via convex hull (scipy), pca / inertia tensor, then dense sampling around major axis (warp).
Besides being off device (and convex hull is expensive) these are often sub-optimal in terms of minimizing volume or surface area. We can brute force over dense samples of all SO(3) and get tighter fits in pure warp.
Oriented bounding boxes are a useful bounding primitive, used for example in Newton.
Newton currently computes them via convex hull (scipy), pca / inertia tensor, then dense sampling around major axis (warp).
Besides being off device (and convex hull is expensive) these are often sub-optimal in terms of minimizing volume or surface area. We can brute force over dense samples of all SO(3) and get tighter fits in pure warp.