**NEW - Structural model - extension on aspect ratio**

### Overview

This experiment extends earlier experiment 05i on the aspect ratio. Because we need to know aspect ratio of training samples we use a little bit modified dataset. So we compute again validation error of previous model 05i.

Choosed values of window aspect ratios are given by statistical measurement from training dataset.

### Results

Parameters of experiment | |
---|---|

imSize [h x w] | 80 x 64 |

winSize [h x w] | 60 x 40 |

window position | half in either axis |

window rotation | {-pi/30; -pi/60 0; pi/60; pi/30} |

window scale | {0.8 1 0.9} |

window ratios ** only for new model | {1.35 1.5 1.65} |

OLD EXPERIMENT 05i | NEW EXPERIMENT 05r | |||||
---|---|---|---|---|---|---|

LAMBDA | Training error | Validation error | nIterations | Training error | Validation error | nIterations |

1 | % | 14.36% | 33 | 4.89% | 12.76% | 36 |

0.1 | 0.02% | 10.72% | 128 | 0.02% | 10.00% | 129 |

**Testing error for minimal validation error: 12.49%**

#### Distribution of aspect ratios through testing dataset

Aspect ratio | Count |
---|---|

1.35 | 643 |

1.5 | 3545 |

1.65 | 905 |

#### Graphical results

In figure 1 you can see twelve random selected samples, detections on whole database are available in file detection.zipFigure 1: Random samples from dataset that haven't basic aspect ratio 1.5 .. also in MATLAB figure

### References

*Choon Hui Teo , S.V.N. Vishwanthan , Alex J. Smola , Quoc V. Le: Bundle Methods for Regularized Risk Minimization, The Journal of Machine Learning Research, 11, p.311-365, 3/1/2010*