Knowing a good feature when you see it: Ground truth and methodology to evaluate local features for recognition
While the majority of computer vision systems are based on representing images by local features, the design of the latter has been so far mostly empirical. In this Chapter we propose to tie the design of local features to their systematic evaluation on a realistic ground-truthed dataset. We propose...
| Main Authors: | , , |
|---|---|
| Format: | Journal article |
| Language: | English |
| Published: |
2010
|
| Summary: | While the majority of computer vision systems are based on representing images by local features, the design of the latter has been so far mostly empirical. In this Chapter we propose to tie the design of local features to their systematic evaluation on a realistic ground-truthed dataset. We propose a novel mathematical characterisation of the co-variance properties of the features which accounts for deviation from the usual idealised image affine (de)formation model. We propose novel metrics to evaluate the features and we show how these can be used to automatically design improved features. © 2010 Springer-Verlag Berlin Heidelberg. |
|---|