Hi,

I have cascaded shadow maps setup and working pretty well in my engine. I have large outdoor scenes and currently only one light (the sun) casting shadows. The problem is that when objects are high up, eg. right above the camera, it is outside the closest, highest detail cascade, and thus does not cast any shadows. My camera usually looks a bit down, since it's a third-person game I am making. How should I go about fixing this? Should I just make the cascades big enough to contain all objects in the scene? This, of course, would mean that the resolution of the cascades goes down.

The high-up objects are mostly static. Should I look into doing some sort of offline precalculated shadow data for these objects?

The same problem also occurs if, say, the camera is looking horizontally forward, the sun is shining on the back of the character and there is a bird flying between the sun and the camera. In that case the bird is outside all of the cascades.

Yes, I am making the cascades as tight as possible (so they cover just what's in the frustum). I have a tweakable value for scaling the cascade near and far clips, but the shadow quality decreases noticeably if I scale them up, presumable because of limited depth buffer resolution, am I right?

EDIT: It's quite possible that I haven't done the CSM correctly, but this is what I am seeing with my current setup. If having infinite cascades is the norm, how do you deal with depth buffer resolution issues?

The "near plane" and "far plane" in my comment above refer to the near and far plane distances of the orthographic projection of the cascade cameras. They have nothing to do with the player camera or where the cascades or shadow casters are.

I can indeed make the cascade "infinite" by moving the camera back by infinity/2 and moving the far plane forward by infinity/2. It's just that earlier when doing this my shadow quality went way down, and I thought this had something to do with depth buffer resolution, but maybe I was wrong.

This is a real problem, though. One (good) solution usually goes by the name "pancaking". Basically, you still set your near and far planes (for the light's ortho projection) to tightly bound the camera frustum. But you make sure not to cull anything behind the near plane (depending on how you do your frustum culling, this is usually pretty easy). Then, in your shadow-mapping vertex shader, after you do your projection, you clamp the Z value:

projPos = mul(matrix_mvp, In.Pos);
projPos.z = max(projPos.z, 0.0f);

This has the effect of flattening everything that's off-screen (from the camera's perspective) onto the near plane of the light's projection. (Because you only really care about the silhouette of those objects). There is the potential for minor artifacts when you have large triangles (they're warped if one vertex is behind the near the plane and another one isn't), but if you just give yourself a small buffer zone (make sure to pull the near plane back a short distance), it should work fine. We're using this technique now, and haven't had any problems with it.

Hmm, very interesting. I need to look into the pancaking technique. Thank you both.