Computer Vision for Rapid Orienting and Classification

Bildverarbeitung für Schnelle Orientierung und Klassifizierung


We build a vision system that analyzes the structure in images [pdf]. It effortlessly provides object proposals with a fast but thorough segmentation. It identifies scenes very well without any pretraining. And it can even categorize decently. For one satellite image database we even exceed the benchmark set by Deep Networks.

     The segmentation process finds any object and scene parts no matter their contrast, no matter their category. This is essential for dealing with novel objects, novel scenes, the anomaly, whereas other approaches (ie. Deep Networks) tend to struggle with unusual situations. The entire output is generated in less than 100 milliseconds for a 480 x 640 pixel image (2.4 GHz) and identifies the straight and curved segments corresponding to the silhouette of objects or scene parts. With that output alone we perform scene identification without any pretraining, as well as coarse classification. When thoroughly parameterized, it can even compete with Deep Networks for classification.

    Our approach is ideal for domestic robots, mobile devices, embedded systems and large image analysis, i.e. satellite images, road scenes, drone images, etc.

Follow on: