Two methods to accurately determine the moment of pupil constriction onset are discussed. For data sampled at a high rate (approximately 200 Hz) pupil velocity deviations from zero can simply be used, giving a satisfactory inaccuracy of about 5 ms. For data sampled at a low rate (less than 50 Hz), e.g. using a TV pupillometer, curve-fitting can be applied. It is demonstrated that curve-fitting, based on a second-order mathematical model, preceded by a linear trend, can result in an inaccuracy of less than 5 ms. Both methods give latencies independent of signal amplitude. This implies that a pupillometer yielding a relative measurement of the pupil area can be used for the detection of pupil constriction latencies. Furthermore it is demonstrated that the averaging of pupil constrictions results in an advanced moment of onset. Where latency differences of less than 25 ms are concerned, raw pupil data should therefore not be averaged.