This paper presents two methods of training complexity reduction by additional selection of features to check in object detector training task by AdaBoost training algorithm. In the first method, the features with weak performance at first weak classifier building process are reduced based on a list of features sorted by minimum weighted error. In the second method the feature similarity measures are used to throw away that features which is similar to earlier checked features with high minimum error rates in possible weak classifiers for current step. Experimental results with MIT-CMU $19\times19$ face detection images show that the error presented by ROC curves is near the same for the learning with and without additional feature reduction during the computational complexity is well reduced.
Authors
Additional information
- DOI
- Digital Object Identifier link open in new tab 10.1007/978-3-319-23814-2_6
- Category
- Aktywność konferencyjna
- Type
- materiały konferencyjne indeksowane w Web of Science
- Language
- angielski
- Publication year
- 2016
