A visual texture descriptor becomes useful for semantic description and classification, if it is highly discriminative and at the same time robust to environmental changes. Features using orientation histograms, such as the widely used SIFT feature, have proven to be very powerful robust local feature descriptors. These features are usually combined in histograms for global characterization. On the other hand, the MFS, which has been defined on simple image features, provides a good global characterization, and has been demonstrated to be robust to a wide range of geometric transformations. In this work the MFS is combined with a powerful feature based on orientation histograms. The resulting descriptor leads to better discriminative power and robustness to both geometric and photometric variations. The approach is applicable to both static and dynamic textures [pdf].
Discretized orientation histogram:
Within windows of different sizes orientation histograms are computed using 8 directions. Then the histograms are quantized based on their topological structure into 29 classes as shown in Fig. 1
Fig. 1.(left): Representation of the orientation histogram. (right): Representative elements for each of the 29 classes of orientation histogram templates.
The complete approach consists of four steps as illustrated in Fig.2.
Fig.2 : Overview of the approach.
Static textures are described by the descriptor given above.
The descriptor for dynamic textures is based on 3D SIFT descriptors, with a 2D histogram that captures the temporal changes.