Difference between eval_attribute and eval_attribute_1 for BSDF visualizations

[xrhan]

Hi, I am implementing a custom integrator that visualize the BSDF parameters (metallic, roughness and specular, etc...) for some computer vision / inverse graphics tasks. I noticed that the two functions eval_attribute and eval_attribute_1 give different outputs in the following test example that I ran (where the 'specular' values are all 0s in eval_attribute).
I wonder which one is the more perceptually / physically accurate to use for encoding the material property of the object? Thank you for your help!

Here is the integrator code:

class BRDFIntegrator(mi.SamplingIntegrator):
    def __init__(self, props):
        super().__init__(props)

    def sample(self, scene, sampler, ray, medium, active):
        with dr.suspend_grad():
            # Detach the incoming ray and compute the surface interaction.
            ray = mi.Ray3f(dr.detach(ray))
            depth = mi.UInt32(0)
            active = mi.Bool(active)
            si = scene.ray_intersect(ray,
                                     ray_flags=mi.RayFlags.All,
                                     coherent=(depth == 0))
            bsdf = si.bsdf(ray)
            prop = "metallic" # "metallic", "specular", "roughness"
            value = bsdf.eval_attribute(prop, si) # eval_attribute, eval_attribute_1

        return (value, si.is_valid(), [])

    def __str__(self):
        return "BRDFParameterIntegrator"

Here is the visualization of the scene BSDF properties, and the integrator results using different functions.

[njroussel]

Hi @xrhan

The short answer is that they are typically identical. I'm not sure why eval_attribute differs from eval_attribute_1 here, is it a standard Mitsuba BSDF ? Which one?

The longer answer is that the behavior can be different if you're using spectral variants (as an example). As mentioned in the documentation the eval_attribute_1 is meant for raw access to the attribute data. Whether that raw data is more "meaningful" depends on your application.