Automatize adjustment of the Shadow Map Projection (by projection the points of interest onto the near clipping plane of the light source)

Hello everyone,

I am new to this so please be forgiving when I cannot explain my situation in a way where everyone instantly understands the problem but I am already thankful for your help!

I am working with OpenGL so Coding in C++.

I have a simple scene where I have one object, a one-point source of light, and a camera. Now I am moving the light around the object and up. When I reached all positions I want, I rotate the object and/or move the camera upwards. When I got all the positions of the light and the camera, I switch the object.
Of course, all the time the lighting and the shadows should be correct. But currently, this is not the case. Which leads me to the…

Sometimes there are shadows in the background where actually shouldn’t be one
[iamge with the wrong shadow in the back][1]

I expect that the shadow map is not big enough to capture this part of the scene. Because of multiple changes in the scene, I want to automate the definition of the size of the shadow map.

What I have so far:
I define the maximal bounding box around every object to know the maximal dimensions here. Through drawing the bounding box and debugging I can say that this is calculated correctly. (I know that I have to take the extent of the base surface into account as well but to clarify if my algorithm works well, I leaf that out for a second)

To geht the extents on the near clipping plane, I project them onto the near clipping plane of the light source to get dimantions for the shadow map I need to use.For the Projection I span the base for the plane with the up and the right vector of the plane. The projection matrix I calculate with A * (A^T * A)^(-1) * A^T
When I now switch the view so that I see the scene from the perspective of the light, I see what I expect: The whole picture is just the object.

[object from the light perspective][2]

So far so good. But when I move the light the to ritht or the left, parts of the image move out of the field of view.
[light moved to the right][3]

When I move the light up the object gets swaged.

[light moved up][4]

Here is my code to adjust the lightsource_projection:

void CGRenderer::adjustLight()
		glm::mat2x3 A(9.0f);
		A[1] = lightsource_camera.getUp();
		A[0] = lightsource_camera.getRight();

		//projection matrix = A(ATA)^-1 AT.
		glm::mat3 P = A * glm::inverse(glm::transpose(A) * A) * glm::transpose(A);
std::vector<glm::vec3> bbPoints;
		bbPoints.push_back(glm::vec3(object.boundingBoxViewSpace[0][0], object.boundingBoxViewSpace[0][1], object.boundingBoxViewSpace[0][2]));
		bbPoints.push_back(glm::vec3(object.boundingBoxViewSpace[1][0], object.boundingBoxViewSpace[0][1], object.boundingBoxViewSpace[0][2]));
		bbPoints.push_back(glm::vec3(object.boundingBoxViewSpace[0][0], object.boundingBoxViewSpace[0][1], object.boundingBoxViewSpace[1][2]));
		bbPoints.push_back(glm::vec3(object.boundingBoxViewSpace[1][0], object.boundingBoxViewSpace[0][1], object.boundingBoxViewSpace[1][2]));
		bbPoints.push_back(glm::vec3(object.boundingBoxViewSpace[0][0], object.boundingBoxViewSpace[1][1], object.boundingBoxViewSpace[0][2]));
		bbPoints.push_back(glm::vec3(object.boundingBoxViewSpace[1][0], object.boundingBoxViewSpace[1][1], object.boundingBoxViewSpace[0][2]));
		bbPoints.push_back(glm::vec3(object.boundingBoxViewSpace[0][0], object.boundingBoxViewSpace[1][1], object.boundingBoxViewSpace[1][2]));
		bbPoints.push_back(glm::vec3(object.boundingBoxViewSpace[1][0], object.boundingBoxViewSpace[1][1], object.boundingBoxViewSpace[1][2]));

		glm::vec3 tmp;
		float x_min = 0.f;
		float x_max = 0.f;
		float y_min = 0.f;
		float y_max = 0.f;
		for (glm::vec3 v : bbPoints)
			tmp = v * P;

			if (tmp.x < x_min)
				x_min = tmp.x;
			if (tmp.x > x_max)
				x_max = tmp.x;
			if (tmp.y < y_min)
				y_min = tmp.y;
			if (tmp.y > y_max)
				y_max = tmp.y;

		parameter.lightprojection_x_min = x_min;
		parameter.lightprojection_x_max = x_max;
		parameter.lightprojection_y_min = y_min;
		parameter.lightprojection_y_max = y_max;

		lightsource_projection = glm::ortho(parameter.lightprojection_x_min, parameter.lightprojection_x_max,
			parameter.lightprojection_y_min, parameter.lightprojection_y_max,
			parameter.lightprojection_z_min, parameter.lightprojection_z_max);

Here my shadow pass function where I adjust the light and use the newly calculated lightsource_projection:

	void CGRenderer::shadowmap_pass()
	glBindFramebuffer(GL_FRAMEBUFFER, framebuffers.shadowmap_buffer);
	glViewport(0, 0, shadowmap_width, shadowmap_height);


	model = glm::scale(glm::mat4_cast(parameter.globalRotation), glm::vec3(parameter.modelScalation));
	model = glm::rotate(glm::mat4(1.f), glm::radians(parameter.modelRotation), glm::vec3(0.f, 1.f, 0.f)) * model;

	glm::mat4 shadow_view = lightsource_camera.getViewMatrix();

	glm::mat4 shadow_model_view = shadow_view * model;


	glUniformMatrix4fv(locs.modelViewProjection, 1, GL_FALSE, glm::value_ptr(lightsource_projection * shadow_model_view));

	glUniformSubroutinesuiv(GL_VERTEX_SHADER, 1, &locs.placementVS);
	glUniformSubroutinesuiv(GL_FRAGMENT_SHADER, 1, &locs.depthmapFS);

	glDrawArrays(GL_TRIANGLES, 0, basesurface.vertsToDraw);

	glDrawArrays(GL_TRIANGLES, 0, object.vertsToDraw);

Why is it not updating the projection correctly? How can I fix that problem? Do I need to change the view matrix to solve the problem?

Any ideas the solve the problem are welcome. If you know different algorithms that do the work, I am happy with that, too.

Kind regards and many thanks!

Cross-ref (for the other pictures):

the definition of your lightsource_camera is missing.
If you translate the lightsource_camera, what it’s pointed toward should slide out of view. Do you miss to change the direction of the lightsource_camera to point toward the model?
After all, the lightsource_camera.getUp() && .getRight() from which you construct P are directions, not positions.

Uh, what? That’s an identity matrix.
A * (AT * A)-1 * AT
= A * (A-1 * (AT)-1) * AT
= (A * A-1) * ((AT)-1 * AT)
= I * I
= I
The actual projection is just that calculated by glm::ortho. Also, this doesn’t seem to be accounting for shadow_model_view (unless object.boundingBoxViewSpace was calculated using the same transformation calculated elsewhere).

Hello GClements,

many thanks for your response!

In theory, you are right. But the matrix A is not necessarily invertible. Especially in my case here because I have a 3x2 matrix.

But (A^T * A) should always be invertible.

Hello CarstenT,

many thanks for your time.

here is the definition of my lightsource_camera:

cgbv::Camera observer_camera, lightsource_camera;
lightsource_camera.setTarget(glm::vec3(0.f, 0.f, 0.f));

Actually the adjustment of the light should not influence the target of the camera. To be sure I call the setTarget() function after every time I call adjustLight(), but I get the same output.

I see. The projection P flattens the points to the plane of the camera.

The flattened box is still in its original coordinate system (presumably view space, not light space), and the ortho projection is based upon that.

Typically, you’d transform the bounding region to the light’s “view” space (i.e. rotation and translation, no projection) and use the transformed region to calculate the bounds for the light’s projection (either orthographic or perspective).

the asymmetry of A (and the further construction) makes me suspicious hold against the asymmetry of the image-error (light-up/down produces a proper orthographic distortion, whereas light-left/right produces a translation)

Here is my solution:

I transform the bounding vertices into the light-view space. Then project it. By using a normalized orthographic projection matrix I directly get the values I want.

Many thanks to @GClements for your help!

	void adjustLight(glm::mat4 shadow_model_view)
	glm::vec4 tmp;
	float x_min = 0.f;
	float x_max = 0.f;
	float y_min = 0.f;
	float y_max = 0.f;
	for (glm::vec3 v : object.boundingVertices) // before bbPoints
		tmp = shadow_model_view * glm::vec4(v, 1.f) ;
		tmp = glm::ortho(-1.f, 1.f, -1.f, 1.f, -1.f, 1.f) * tmp;

		if (tmp.x < x_min)
			x_min = tmp.x;
		if (tmp.x > x_max)
			x_max = tmp.x;
		if (tmp.y < y_min)
			y_min = tmp.y;
		if (tmp.y > y_max)
			y_max = tmp.y;
            // z_min und z_max are predefined
	lightsource_projection = glm::ortho(x_min, x_max, y_min, y_max, z_min, z_max);

PS: @CarstenT you were richt as well. In the construction of A the location of the plane was missing. So in the end I had a 3x3 matrix but it still did not work.

A rare treat to have the suggestions crystalise into a presented solution. Thanks