The paper propose an algorithm to train an object detection model, which is very fast and its detection power is competitive to the state-of-art at then. There are three main contribution of their works:
1. Integral Image
The complexity of computing the integral image is low - O(image size). And more importantly, it can be used to obtain any haar-like feature in constant time. This speeds up the detection process. Rare features can be extracted as fast as it.
Integral image has another usage. We can get the variance within any rectangle in the image in constant time if we have to integral image, with one accummulating the intensity value and the other accumulating the square of intensity value.
2. Training by Adaboost
There are tons of haar-like feature in an image, but many of those features are useless. By treating each weak classifier as an indicator of the usefulness of a haar-like feature, AdaBoost can be used to select a set of features and train the classifier simutaneously.
Since feature selection is done automatically, their work can be easily extended to objects other than faces.
3. Attentional Cascade
This part is designed for accelerating the detection speed rather than the accuracy. The idea is simple - sacrificing the false positive rate to achieve very high detection rate in the beginning to thrown out as many unlikely candidate region as possible. Since most candidate regions are removed in the beginning, the number of operations in the latter stage is decreased.
I think this is a very classic work in object detect. It can be used to detect objects other than faces without any prior knowledge as long as you have enough training data. Of course, it doesn't quarantee to perform well on all objects. Haar-like feature is a kind of appearance feature. If objects' appearance often change (e.g. t-shirt), the method proposed in the paper would easily fail.
Reference:
P. Viola and M. Jones, "Rapid object detection using a boosted cascade of simple features." Proc. CVPR, Vol. 1, pp. 511-518, 2001.
1. Integral Image
The complexity of computing the integral image is low - O(image size). And more importantly, it can be used to obtain any haar-like feature in constant time. This speeds up the detection process. Rare features can be extracted as fast as it.
Integral image has another usage. We can get the variance within any rectangle in the image in constant time if we have to integral image, with one accummulating the intensity value and the other accumulating the square of intensity value.
2. Training by Adaboost
There are tons of haar-like feature in an image, but many of those features are useless. By treating each weak classifier as an indicator of the usefulness of a haar-like feature, AdaBoost can be used to select a set of features and train the classifier simutaneously.
Since feature selection is done automatically, their work can be easily extended to objects other than faces.
3. Attentional Cascade
This part is designed for accelerating the detection speed rather than the accuracy. The idea is simple - sacrificing the false positive rate to achieve very high detection rate in the beginning to thrown out as many unlikely candidate region as possible. Since most candidate regions are removed in the beginning, the number of operations in the latter stage is decreased.
I think this is a very classic work in object detect. It can be used to detect objects other than faces without any prior knowledge as long as you have enough training data. Of course, it doesn't quarantee to perform well on all objects. Haar-like feature is a kind of appearance feature. If objects' appearance often change (e.g. t-shirt), the method proposed in the paper would easily fail.
Reference:
P. Viola and M. Jones, "Rapid object detection using a boosted cascade of simple features." Proc. CVPR, Vol. 1, pp. 511-518, 2001.
沒有留言:
張貼留言