In this paper we describe and discuss our existing PicSOM software framework from the point of view of context-adaptive analysis of image contents, especially its method for using automatic image segmentation. We describe and experimentally validate a modification to the segment-using procedure that both essentially reduces the computational cost and slightly improves classification accuracy. Finally, we apply the segment-using methodology in qualitatively investigating the roles of primary objects and their context in classifying the images of the Pascal VOC Challenge 2006 database.