Syseca and IGN are working on various steps in the ongoing march from digital photogrammetry to the semi-automation and ultimately the full automation of data manipulation, i.e., capture and analysis. The immediate goals are to reduce the production costs and the data availability delays. Within this context, we have tackle the distinctive problem of 'automated road network extraction.' The methodology adopted is to first study semi-automatic solutions which probably increase the global efficiency of human operators in topographic data capture; in a second step, automatic solutions are designed based upon the gained experience. We report on different (semi-)automatic solutions for the road following algorithm. One key aspect of our method is to have the stages of 'detection' and 'geometric recovery' cooperate together while remaining distinct. 'Detection' is based on a local (texture) analysis of the image, while 'geometric recovery' is concerned with the extraction of 'road objects' for both monocular and stereo information. 'Detection' is a low-level visual process, 'reasoning' directly at the level of image intensities, while the mid-level visual process, 'geometric recovery', uses contextual knowledge about roads, both generic, e.g. parallelism of borders, and specific, e.g. using previously extracted road segments and disparities. We then pursue our 'march' by reporting on steps we are exploring toward full automation. We have in particular made attempts at tackling the automation of the initialization step to start searching in a valid direction.