Weak precision when reading framebuffer with glReadPixels(...)

Question

I want to leverage OpenGL to sample a discrete scalar map. The scalar map is given on per-triangle basis, i.e. one scalar per triangle.

First, I draw each triangle with flat shading and a color value corresponding to the scalar map. Then, I read the sampled scalar map from the framebuffer using glReadPixels(...).

My problem is as follows:

For some reason I can only recover color values up to 1e-2! I have no clue why this precision is so bad. I'm writing and reading floats from the framebuffer.

Does anyone know why I can recover only such a bad floating point precision from the framebuffer?

Here is the important code snippet, where I simply use a constant color value colorValue for each triangle. If I set colorValue e.g. to 1e-4, I can only recover zeros from the framebuffer.

// Set viewport to desired map size
glViewport(0,0,imageWidth, imageHeight);
glClearColor(0.,0.,0.,1.); // Black background
glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT); 

glDisable(GL_LIGHTING);
glMatrixMode(GL_PROJECTION);
glLoadIdentity();
gluOrtho2D(-1,1.,-1,1);
glMatrixMode(GL_MODELVIEW);

// Draw scalar map
glBegin(GL_TRIANGLES);
for (unsigned int f=0; f<F.rows(); ++f)
 {
  val = colorVal;
  glColor3f(val, val, val);

  glVertex3f(...);
  glVertex3f(...);
  glVertex3f(...);
}
glEnd();
glFlush();
glFinish();

GLfloat* fbuffer = new GLfloat[imageWidth*imageHeight];
glReadPixels(0,0,imageWidth,imageHeight,GL_RED,GL_FLOAT,fbuffer);

How are you creating your GL Buffers? Sounds like you're not creating a true FP buffer. — Yeraze, May 19 '14 at 21:26

Andon M. Coleman · Accepted Answer · 2014-05-19T21:34:33.553

It does not matter what you are writing to or reading from the framebuffer, so much as it does what the framebuffer actually stores.

GL will happily convert colors to/from unsigned normalized all day long. Unless you have a floating-point framebuffer, lots of unpleasant things happen such as clamping ([0.0,1.0]) and scaling into the range of fixed-point values (8-bit usually).

Given an 8-bit unsigned normalized colorbuffer, the smallest non-zero color value you can store and recover is ~0.0039 (1.0/255.0).

The process of creating a floating-point default framebuffer is platform/framework specific. To do what you want portably, the easiest thing is actually to draw into an FBO with a floating-point color attachment that you create after you setup the default framebuffer.

Thanks for pointing in the correct direction! I'm quite new to OpenGL so I did not know about FBOs. The answer posted [here](http://stackoverflow.com/questions/10950743/low-precision-and-clamping-when-writing-to-floating-point-fbo) solved my problem. Thanks a lot! — user1086105, May 20 '14 at 11:59

Weak precision when reading framebuffer with glReadPixels(...)

1 Answers1