The retinal flow during normal locomotion contains components due to rotation and translation of the observer. The translatory part of the flow-pattern is informative of heading, because it radiates outward from the direction of heading. However, it is not directly accessible from the retinal flow. Nevertheless, humans can perceive their direction of heading from the compound retinal flow without need for extra-retinal signals that indicate the rotation. Two classes of models have been proposed to explain the visual decomposition of the retinal flow into its constituent parts. One type relies on local operations to remove the rotational part of the flow field. The other type explicitly determines the direction and magnitude of the rotation from the global retinal flow, for subsequent removal. According to the former model, nearby points are most reliable for estimating one's heading. In the latter type of model the quality of the heading estimate depends on the accuracy with which the ego-rotation is determined and is therefore most reliable when based on the most distant points. We report that subjects underestimate the eccentricity of heading, relative to the fixated point in the ground plane, when the visible range of the ground plane is reduced. Moreover we find that in perception of heading, humans can tolerate more noise than the optimal observer (in the least squares sense) would do if only using optic flow. The latter finding argues against both schemes because ultimately both classes of model are limited in their noise tolerance to that of the optimal observer, which uses all information available in the optic flow. Apparently humans use more information than is present in the optic flow. Both aspects of human performance are consistent with the use of static depth information in addition to the optic flow to select the most distant points. Processing of the flow of these selected points provides the most reliable estimate of the ego-rotation. Subsequent estimates of the heading direction, obtained from the translatory component of the flow, are robust with respect to noise. In such a scheme heading estimates are subject to systematic errors, similar to those reported, if the most distant points are not much further away than the fixation point, because the ego-rotation is underestimated.