Move camera to fit 3D scene

Question

Hi there.

I'm looking for an algorithm to fit a bounding box inside a viewport (in my case a DirectX scene). I know about algorithms for centering a bounding sphere in a orthographic camera but would need the same for a bounding box and a perspective camera. I can not just change the FOV because this app has FOV as a user editable variable, so it must move the camera.

I have most of the data:

• I have the up-vector for the camera
• I have the center point of the bounding box
• I have the look-at vector (direction and distance) from the camera point to the box center
• I have projected the points on a plane perpendicular to the camera and retrieved the coefficients describing how much the max/min X and Y coords are within or outside the viewing plane.

Problems I have:

• Center of the bounding box isn't necessarily in the center of the viewport (that is, it's bounding rectangle after projection).
• Since the field of view "skew" the projection (see http://en.wikipedia.org/wiki/File:Perspective-foreshortening.svg) I cannot simply use the coefficients as a scale factor to move the camera because it will overshoot/undershoot the desired camera position

How do I find the camera position so that it fills the viewport as pixel perfect as possible (exception being if the aspect ratio is far from 1.0, it only needs to fill one of the screen axis)?

I've tried some other things:

• Using a bounding sphere and Tangent to find a scale factor to move the camera. This doesn't work well, because, it doesn't take into account the perspective projection, and secondly spheres are bad bounding volumes for my use because I have a lot of flat and long geometries.
• Iterating calls to the function to get a smaller and smaller error in the camera position. This has worked somewhat, but I can sometimes run into weird edge cases where the camera position overshoots too much and the error factor increases. Also, when doing this I didn't recenter the model based on the position of the bounding rectangle. I couldn't find a solid, robust way to do that reliably.

Help please!

Answer 1

There are many possible camera positions + orientations where the bounding box would fit inside the view frustum. But any procedure would select one specific camera position and orientation.

If you would consider bounding spheres, one solution could be to

• first change orientation to look at bounding sphere center
• then move back sufficiently (negative look direction) for bounding sphere to fit inside frustum

With bounding boxes you could consider an earlier step of first positioning the camera at perpendicular to the center of the largest (or smallest, whatever you prefer) cube face.

I have no experience with DirectX, but moving and changing the looking direction of the camera to center a certain point should be easy. The hard part is to do the math of deciding how far to move to view the object.

Math

If you know the bounding size s of the object in world coordinates (we are not interested in pixels or camera coordinates, since those are dependent on your distance) from the orientation of the camera, you can compute the required distance d of the camera to the bounding shape if you know the x and y Field-Of-View angle a of the perspective projection.

     frustum      ------            
            ------    *****          -  
       -----          *   *          |
   -===     ) FOV a   *bounding box  | BB size s
camera -----          *   *          |
            ------    *****          -
                  ------

  |-------------------|
        distance d

So, the math is tan(a/2) = (s/2) / d => d = (s/2) / tan(a/2) Which will give you the distance the camera should be placed from the closest bounding surface.

Answer 2

Since you have a bounding box, you should have a basis describing it's orientation. It seems that you want to position the camera on the line coincident with the basis vector describing the smallest dimension of the box, then roll the camera so that the largest dimension is horizontal (assuming you have OBB and not AABB). This assumes that the aspect ratio is greater than 1.0; if not you'll want to use the vertical dimension.

What I would attempt:

1. Find the smallest box dimension.
2. Find the associated basis vector.
3. Scale the basis vector by the distance from the center of the box the camera should be. This distance is just boxWidth / (2 * tan(horizontalFov / 2)). Note that boxWidth is the width of the largest dimension of the box.
4. Place the camera at boxCenter + scaledBasis looking at the boxCenter.
5. Roll the camera if necessary to align the camera's up vector with the appropriate box basis vector.

Edit:

So I think what you're getting at is that you have the camera at an arbitrary position looking somewhere, and you have an AABB at another position. Without moving the camera to face a side of the box, you want to:

• Look at the center of the box
• Translate the camera along it's look vector so that the box takes the maximum amount of screen space

If this is the case you'll have a bit more work; here's what I suggest:

1. Rotate the camera to look at the center of the bounding box.
2. Project all the points of the box into screen space and find the min/max bounding box in screen space (you already have this).
3. Now Unproject two opposing corners of the screen space bounding box into world space. For a Z value use the closest world space points of your AABB to the camera.
4. This should get you a world space plane facing the camera positioned at the point on the AABB that is closest to the camera.
5. Now use our existing side-facing method to move the camera to the appropriate spot, treating this plane as the side of your box.

Answer 3

I don't have it at hand at the moment but the book you want is http://www.amazon.com/Jim-Blinns-Corner-Graphics-Pipeline/dp/1558603875/ref=ntt_at_ep_dpi_1

He has a whole chapter on this

Answer 4

This is copied straight from my engine, it creates 6 planes which represent each of the six sides of the frutsum. I hope it comes in useful.

internal class BoundingFrustum
    {
        private readonly float4x4 matrix = new float4x4(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
        private readonly Plane[] planes;

        internal BoundingFrustum(float4x4 value)
        {
            planes = new Plane[6];
            for (int i = 0; i < 6; i++)
                planes[i] = new Plane();
            Setfloat4x4(value);
        }

        private void Setfloat4x4(float4x4 value)
        {
            planes[2].Normal.X = -value.M14 - value.M11;
            planes[2].Normal.Y = -value.M24 - value.M21;
            planes[2].Normal.Z = -value.M34 - value.M31;
            planes[2].D = -value.M44 - value.M41;
            planes[3].Normal.X = -value.M14 + value.M11;
            planes[3].Normal.Y = -value.M24 + value.M21;
            planes[3].Normal.Z = -value.M34 + value.M31;
            planes[3].D = -value.M44 + value.M41;
            planes[4].Normal.X = -value.M14 + value.M12;
            planes[4].Normal.Y = -value.M24 + value.M22;
            planes[4].Normal.Z = -value.M34 + value.M32;
            planes[4].D = -value.M44 + value.M42;
            planes[5].Normal.X = -value.M14 - value.M12;
            planes[5].Normal.Y = -value.M24 - value.M22;
            planes[5].Normal.Z = -value.M34 - value.M32;
            planes[5].D = -value.M44 - value.M42;
            planes[0].Normal.X = -value.M13;
            planes[0].Normal.Y = -value.M23;
            planes[0].Normal.Z = -value.M33;
            planes[0].D = -value.M43;
            planes[1].Normal.X = -value.M14 + value.M13;
            planes[1].Normal.Y = -value.M24 + value.M23;
            planes[1].Normal.Z = -value.M34 + value.M33;
            planes[1].D = -value.M44 + value.M43;
            for (int i = 0; i < 6; i++)
            {
                float num2 = planes[i].Normal.Length();
                planes[i].Normal = planes[i].Normal / num2;
                planes[i].D /= num2;
            }
        }

        internal Plane Bottom
        {
            get { return planes[5]; }
        }
        internal Plane Far
        {
            get { return planes[1]; }
        }
        internal Plane Left
        {
            get { return planes[2]; }
        }
        internal Plane Near
        {
            get { return planes[0]; }
        }
        internal Plane Right
        {
            get { return planes[3]; }
        }
        internal Plane Top
        {
            get { return planes[4]; }
        }
    }