I have a bounding box class with a min (x,y,z) and max point (x,y,z). The object that is being rotated has a rotation matrix (3x3 matrix), but I don't know how to update the bounding box with rotations.

I'm also a little confused if I should be using Oriented Bounding Boxes instead. For example, if the 3D object is a cube then the bounding box would rotate with it so they share the same 8 vertices at whatever rotation the cube is in.

What's the best way to represent the bounding box for rotations and test for collisions?

  • \$\begingroup\$ May I ask why you deleted your previous question on a similar topic? With regard to this question, the choice of whether to use axis aligned bounding boxes or oriented bounding boxes is a judgement call you'll need to make based on your application. Is speed of rejection important enough to tolerate some extra false positives (eg. for quickly thinning the set of colliding objects in the first broadphase pass of a collision detection routine), or do you need more accurate estimates of rotated objects, enough to justify the more complicated checks (eg. tightly fitting a costly graphical effect)? \$\endgroup\$ – DMGregory Aug 15 '18 at 1:08
  • \$\begingroup\$ I deleted my other question, because while right now my methods are primarily based on AABB, I really don't know how to recalculate the min/max from rotations so it continues to work with my collision check functions. I thought maybe I was doing it wrong with AABB, and thought OBB might be better for handling rotations. Like I said in the former question, I have a rotation matrix, but don't know how to use that data to create the right boundaries for collision checks. \$\endgroup\$ – karamazovbros Aug 15 '18 at 1:12
  • \$\begingroup\$ As reference I'm using most of this code for checking collisions with AABB: developer.mozilla.org/en-US/docs/Games/Techniques/… \$\endgroup\$ – karamazovbros Aug 15 '18 at 1:12

Given an original axis-aligned bounding box and a transformation matrix, you can compute a new axis-aligned bounding box from the extents of the original box, transformed by the matrix, like so:

// Keep a scratchpad array we can re-use for listing all 8 corners of a box.
Vector3[] corners = new Vector3[8];

// Fill our scratchpad array with the corners of an axis-aligned bounding box.  
void PopulateCorners(AABB bounds) {
    corners[0] = bounds.Min;
    corners[1] = new Vector3(bounds.Min.x, bounds.Min.y, bounds.Max.z);
    corners[2] = new Vector3(bounds.Min.x, bounds.Max.y, bounds.Min.z);
    corners[3] = new Vector3(bounds.Max.x, bounds.Min.y, bounds.Min.z);
    corners[4] = new Vector3(bounds.Min.x, bounds.Max.y, bounds.Max.z);
    corners[5] = new Vector3(bounds.Max.x, bounds.Min.y, bounds.Max.z);
    corners[6] = new Vector3(bounds.Max.x, bounds.Max.y, bounds.Min.z);
    corners[7] = bounds.Max;

// Compute a new axis-aligned bounding box that will contain whatever the original
// bounds did, after an affine transformation. (Note this is a lossy operation)
AABB GetTransformedBounds(AABB original, Matrix transformation) {

    Vector3 min = new Vector3(1, 1, 1) * float.PositiveInfinity;
    Vector3 max = new Vector3(1, 1, 1) * float.NegativeInfinity;

    // Transform all of the corners, and keep track of the greatest and least
    // values we see on each coordinate axis.
    for(int i = 0; i < 8; i++) {
        Vector3 transformed = transformation * corners[i];
        min = Vector3.ComponentwiseMin(min, transformed);
        max = Vector3.ComponentwiseMax(max, transformed);

    return new AABB(min, max);

Note that you'll want to save the object's original "neutral orientation" bounding box separately. Anytime you want to compute a set of rotated bounds for the object, always work from this original, not from a previous rotated result. The reason is that an AABB might not be a tight fit - it might leave some room at the corners for example. If we rotate that box, then take the axis-aligned bounds of the rotated box, we'll tend to accumulate extra padding around the object, and our bounds will become less and less tight. Working always from the original ensures this error can't accumulate, and we get the best fit we can under the circumstances.

Of course, you can be more accurate still by always working from the source object/mesh/primitives. Transform all the points of the object (or its convex hull for simplicity) and then compute a new min & max from scratch off of those values. That will get the absolute tightest axis-aligned bounding box, like the one shown in the gif of the rotating knot in the article you link. But it's also often more work than just transforming the AABB.


Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Not the answer you're looking for? Browse other questions tagged or ask your own question.