Final Report Summary - LEARNTEX (Learning Texture Descriptors)
In this project, we aimed i) to develop new approaches for learning descriptors for textures and texture-like objects, and ii) to establish informative benchmarking standards for evaluating texture classification.
The desiderata for the first goal were good performance (ability to recognise and synthesize), scalability (efficient representation and recognition) and effectiveness (learning new patterns from small number of examples). We have designed descriptors which are hierarchical, efficient, and notably, rotation-invariant.
The approach has been tested on texture-like objects. In particular, it has been applied to detection of windows in facades, which is an important case of texture-like object detection in practice.
As for the establishing the benchmarking standards, we have shown that an evaluation methodology for benchmarking texture classification which has been in use for around 10 years was only very little informative. This fact resulted from that there was not sufficient separation of the training and the test data, and thus the classifiers could easily overfit, without incurring any penalty. We have corrected the methodology by properly separating the data, and published the corrected methodology in a form which enables the computer vision community to use it.
The impact of the results of this project is again two-fold.
The research in the descriptors for textures and texture-like objects will continue, and the results obtained so far have the potential to contribute to the important practical tasks (e. g. Semantic description of street images). The evaluation methodology enables us to compare the existing algorithms properly, and to correctly evaluate any future methods for texture classification.
The desiderata for the first goal were good performance (ability to recognise and synthesize), scalability (efficient representation and recognition) and effectiveness (learning new patterns from small number of examples). We have designed descriptors which are hierarchical, efficient, and notably, rotation-invariant.
The approach has been tested on texture-like objects. In particular, it has been applied to detection of windows in facades, which is an important case of texture-like object detection in practice.
As for the establishing the benchmarking standards, we have shown that an evaluation methodology for benchmarking texture classification which has been in use for around 10 years was only very little informative. This fact resulted from that there was not sufficient separation of the training and the test data, and thus the classifiers could easily overfit, without incurring any penalty. We have corrected the methodology by properly separating the data, and published the corrected methodology in a form which enables the computer vision community to use it.
The impact of the results of this project is again two-fold.
The research in the descriptors for textures and texture-like objects will continue, and the results obtained so far have the potential to contribute to the important practical tasks (e. g. Semantic description of street images). The evaluation methodology enables us to compare the existing algorithms properly, and to correctly evaluate any future methods for texture classification.