We present a method for interactive global illumination computation which is embedded in the framework of

Quasi-Monte Carlo photon tracing and density estimation techniques. The method exploits temporal coherence of

illumination by tracing photons selectively to the scene regions that require illumination update. Such regions are

identified with a high probability by a small number of the pilot photons. Based on the pilot photons which require

updating, the remaining photons with similar paths in the scene can be found immediately. This becomes possible

due to the periodicity property inherent to the multi-dimensional Halton sequence, which is used to generate

photons. If invalid photons cannot all be updated during a single frame, frames are progressively refined in subsequent

cycles. The order in which the photons are updated is decided by inexpensive energy- and perception-based

criteria whose goal is to minimize the perceivability of outdated illumination. The method buckets all photons

on-the-fly in mesh elements and does not require any data structures in the temporal domain, which makes it suit-able

for interactive rendering of complex scenes. Since mesh-based reconstruction of lighting patterns with high

spatial frequencies is inefficient, we use a hybrid approach in which direct illumination and resulting shadows are

rendered using graphics hardware.