Good question. I never saw the bias on a per-object basis. The bias depends among other factors on the curvature of the surface, the size of the projected texels and also the radius of the light source (depth range). So I think it is not possible to find a global per-object bias that would work with all lights. But what I have been thinking about is having a different bias for each shadow map split (for the cascaded SMs).
We had the same problem as you back in the University days with our huge city scene. We found a solution that clearly reduces the biasing problem but requires ShaderModel 3 hardware. It is just a single line; it is already in the shaders but commented out. Look at the SHADOWMAP contexts.
// Clearly better bias but requires SM 3.0
gl_FragDepth = dist + abs( dFdx( dist ) ) + abs( dFdy( dist ) ) + shadowBias;
In case you wonder about the dFdx/dFdy functions: They describe the rate of change of a value relative to the neighboring pixels. So this method better takes into account the curvature of surfaces.