Computer-rendered HDR and LDR 4k images database

Realistic image computation mimics the natural process of acquiring pictures by simulating the physical interactions of light between all the objects, lights and cameras lying within a modelled 3D scene. This process is known as global illumination and was formalised by Kajiya with the following rendering Equation:
\(\begin{equation} \label{eq:rendering_equation} L_o(x, \omega_o) = {L_e(x, \omega_o)} + \int_{\Omega}^{} {L_i(x, \omega_i)} \cdot f_r(x, \omega_i \rightarrow \omega_o) \cdot \cos \theta_i d\omega_i \end{equation}\)

where:

 \(L_o(x, \omega_o)\) is the luminance traveling from point \(x\) in direction \(\omega_o\);

\(L_e(x, \omega_o)\) is point \(x\) emitted luminance (it is null if point x does not lie on a ligth source surface);

the integral represents the set of luminances \(L_i\)incident in \(x \) from the hemisphere of the directions \(\Omega\) and reflected in the direction \(\omega_o\). The reflected luminances are weighted by the materials reflecting properties (bidirectionnal reflectance function \(f_r(x, \omega_i \rightarrow \omega_o)\)) and the cosinus of the incident angle.

This equation cannot be analytically solved and Monte Carlo approaches are generally used to estimate the value of the pixels of the final image.

This proposed dataset is composed of 32 points of view of photo realistics images with different level of samples (following the Monte Carlo approach) for each. Each image is 3840 × 2160 pixels in size. The most noisy image is of 2⁰ samples and the reference one (the most converged image obtained) is of 2²⁰ samples. The pbrt rendering engine (version 4) was used to generate these images.