It's been some time since I've worked with this type of processing, so forgive me if some of these answers are wrong.

1.  Each selected feature will have a specific (relative) position.  This is done so that, in a face for example, you're likely measuring the feature from the same part of the face.

2. Features can be selected more than once.  I think it is unlikely that the same feature will be selected more than once in the same stage but I think it is technically possible.

1. No, the Haar features are used in a sliding window manner, just like filtering or convolution. Therefore, the Haar kernels are applied to every single pixel. If the feature was only applied to the center of the image, the face detection algorithm wouldn't work (unless the face was always in the center - but then you probably wouldn't need a face detection algorithm). This is one of the reasons that the Haar wavelets are used, because the integral image representation allows them to be calculated with many fewer computations than standard convolution.

2. Features can be used more than once. In my own work, I noticed that this happens quite a bit, but in reality they are not exactly the same features, because along with each Haar wavelet, there is a threshold value that converts the Haar output to a binary decision, and typically when a feature is re-used, it is re-used with a different threshold value.

Charles is correct about point 1, I misunderstood the question.  If you have a large image, the face detector is run in a sliding-window way so that nearly every possible position of a face in the larger image is tested.  However, within each individual window, each feature has a specific location.

1. Haar features are used like convolution window. The resultant feature set will be given to the classifier. we cant predict a a specific position of the face in a big image.

2. features can be used more than once but with different threshold values.

 For the question 1, I think what you should keep in mind is that, classical Haar filters are sensitive to position changes. So, for the evaluated windows, those Haar filters selected in training stages should be placed in the same positions with that in training images. However, because the sliding windows tend to have different scales from training images, so you need to 'align' the sliding window with training scale via some techniques before you place those Haar filters.

Hi all, thanks for the answers ! I understand that the face may not be in the center of the image.

So just to clarify with reference to the attached image, the face detector is run in a sliding-window manner and, at all instances, the features will be placed in a specific position in the sub-window for evaluation. At the end, the window and feature will then be re-scaled and the entire procedure repeated.  I understand because the haar features are used in a sliding window manner, there will be many more "positions" between "A" and "B" in the attached image.

Would this be the correct explanation ?

Again thanks for all the inputs :)

I think the step 'Rescale' operation should be conducted before we compute Haar features, because the specific positions where we place the Haar filters should strictly refer to the window scale.

For the rescale problem, in my experience, two ways we can choose to compute Haar features for the evaluated windows:

1. Resize the window image. Resize the evaluated window to the same scale with training images, then you can easily place Haar filters in the specific positions to compute Haar features;

2. Resize Haar filters. Resize the Haar filters (Rv times) and keep the aspect ratio between resized Haar filters and evaluated window as same as the ratio between original Haar filters and training images. The Haar features computed in this way should be further divided by the times Rv, because the Haar area is changed as the filter scale changes.

The solution 1 is easy to implement, but it's a bit time-consuming in the sliding window manner as the image resize operation is slow.