Problem with Reconstructing Position From Depth Buffer

After trying several methods that all had some sort of limitation or problem I found a method of reconstructing world positions from a depth buffer that works but is only accurate when the camera x and z rotation is set to zero, but it's the most flexible method I've found so far (can be applied to non-post processing shaders). Here's the reconstruction function:

Code (CSharp):

1. float3 DepthToWorld(float2 screenPos, float depth)
2.             {
3.                 float3 cameraPos = _WorldSpaceCameraPos;
4.                 float3 localX = _FrustumCorners[1] - _FrustumCorners[0];
5.                 float3 localY = _FrustumCorners[3] - _FrustumCorners[0];
6.                
7.                 float3 ray = screenPos.x * localX + screenPos.y * localY + _FrustumCorners[0];
8.                
9.                 float3 pos = cameraPos + (-ray) * depth;
10.                 return pos;
11.             }

and the code that calls it:

Code (CSharp):

1. float3 worldPos = DepthToWorld(i.uv, depth);

and the code that creates the frustum corners matrix:

Code (CSharp):

1.         float CAMERA_NEAR = camera.nearClipPlane;
2.         float CAMERA_FAR = camera.farClipPlane;
3.         float CAMERA_FOV = camera.fieldOfView;
4.         float CAMERA_ASPECT_RATIO = camera.aspect;
5.        
6.         Matrix4x4 frustumCorners = Matrix4x4.identity;      
7.        
8.         float fovWHalf = CAMERA_FOV * 0.5f;
9.        
10.         Vector3 toRight = camera.transform.right * CAMERA_NEAR * Mathf.Tan (fovWHalf * Mathf.Deg2Rad) * CAMERA_ASPECT_RATIO;
11.         Vector3 toTop = camera.transform.up * CAMERA_NEAR * Mathf.Tan (fovWHalf * Mathf.Deg2Rad);
12.        
13.         Vector3 topLeft = (camera.transform.forward * CAMERA_NEAR - toRight + toTop);
14.         float CAMERA_SCALE = topLeft.magnitude * CAMERA_FAR/CAMERA_NEAR;  
15.        
16.         topLeft.Normalize();
17.         topLeft *= CAMERA_SCALE;
18.        
19.         Vector3 topRight = (camera.transform.forward * CAMERA_NEAR + toRight + toTop);
20.         topRight.Normalize();
21.         topRight *= CAMERA_SCALE;
22.        
23.         Vector3 bottomRight = (camera.transform.forward * CAMERA_NEAR + toRight - toTop);
24.         bottomRight.Normalize();
25.         bottomRight *= CAMERA_SCALE;
26.        
27.         Vector3 bottomLeft = (camera.transform.forward * CAMERA_NEAR - toRight - toTop);
28.         bottomLeft.Normalize();
29.         bottomLeft *= CAMERA_SCALE;
30.  
31.         frustumCorners.SetRow (0, topLeft);
32.         frustumCorners.SetRow (1, topRight);      
33.         frustumCorners.SetRow (2, bottomRight);
34.         frustumCorners.SetRow (3, bottomLeft);
35.  
36.         mat.SetMatrix ("_FrustumCorners", frustumCorners);

and where I originally found this method: https://www.synapsegaming.com/forums/p/1165/13827.aspx (the last post on the page). As I said it seems to work really well but when the camera rotates on the x or z axis the reconstruction doesn't account for the rotation (but does for the y axis). Does anyone what could be going wrong?

 

Am I supposed to be doing a simple matrix multiplication somewhere? Everything seems to check out so I wouldn't think that it would be necessary.

 

Nevermind I figured it out, a few simple errors, the ray variable should not be negative in the line:

Code (CSharp):

1. float3 pos = cameraPos + (-ray) * depth;

and the y component of screenPos in the line:

Code (CSharp):

1. float3 ray = screenPos.x * localX + screenPos.y * localY + _FrustumCorners[0];

has to be inverted and the line should read:

Code (CSharp):

1. float3 ray = screenPos.x * localX + (1-screenPos.y) * localY + _FrustumCorners[0];

and now it works perfectly.