This paper describes a highly successful two-stage segmentation technique. The performance of this technique has been quantified on a large set of complex, outdoor scenes. In the first stage, a self-organising feature map is trained to perform the segmentation task. The addition of colour information, and the quantification of the technique using real-world data, extends the work of Mao and Jain. The second stage of the segmentation process overcomes the problem of over-segmentation by merging regions. It achieves this through use of a multi-layer perceptron.