Object Type Classification Using Structure-based Feature Representation
Tomoyuki Nagahashi, Hironobu Fujiyoshi, Takeo Kanade · 2007
Current feature-based object type classification meth-ods information of texture and shape based informa-tion derived from image patches. Generally, input fea-tures, such as the aspect ratio, are derived from rough characteristics of the entire object. However, we de-rive input features from a parts-based representation of the object. We propose a method to distinguish ob-ject types using structure-based features described by a Gaussian mixture model. This approach uses Gaus-sian fitting onto foreground pixels detected by back-ground subtraction to segment an image patch into sev-eral sub-regions, each of which is related to a physical part of the object. The object is modeled as a graph, where the nodes contain SIFT(Scale Invariant Feature Transform) information obtained from the correspond-ing segmented regions, and the edges contain informa-tion on distance between two connected regions. By calculating the distance between the reference and input graphs, we can use a k-NN-based classifier to classify an object as one of the following: single human, hu-man group, bike, or vehicle. We demonstrate that we can obtain higher classification performance when us-ing both conventional and structure-based features to-gether than when using either alone. 1.