In this paper we propose a camera-based three-stage approach for the automatic layout analysis of whiteboard documents. Assuming a reasonable foreground-background separation of the handwriting it starts with a locally adaptive binarization followed by connected component extraction. The latter are then automatically classified as representing either simple graphical elements of a mindmap or elementary text patches. In the final stage the text patches are subject to a clustering procedure in order to generate hypotheses for those image regions where textual annotations of the mindmap can be found.

In order to demonstrate the effectiveness of the proposed approach we report results of a writer independent experimental evaluation on a data set of mindmap images created by several different writers without any constraints on writing or drawing style.