Standard density estimation approaches suffer from visible bias due to low-pass filtering of the lighting function. Therefore, most photon density estimation methods have been used primarily with inefficient Monte Carlo final gathering to achieve high-quality results for the indirect illumination. We present a density estimation technique for efficiently computing all-frequency global illumination in diffuse and moderately glossy scenes. In particular, we compute the direct, indirect, and caustics illumination during photon tracing from the light sources. Since the high frequencies in the illumination often arise from visibility changes and surface normal variations, we consider a kernel that takes these factors into account. To efficiently detect visibility changes, we introduce a hierarchical voxel data structure of the scene geometry, which is generated on GPU. Further, we preserve the surface orientation by computing the density estimation in ray space.