The author observes that the object has a very good contour when it is given to a small scale.
Of That is, the edge gradient of the target is more obvious, and the combination becomes a closed contour. The goal here is to be generalized,
Can be any type of object. ( basis for the conclusion of the thesis )
Figure A. Represents the original image, and Figure B represents a gradient image,
Then the author zooms to a lot of scales, figure C, is
After zooming the gradient image to 8x8, the positive and negative
example, Figure D shows a 64-bit vector as the
feature, a right that is obtained using linear SVM training
The weight factor graph. Where the red box in Figure A indicates
Target, the green box indicates a non-target
Feature Extraction Bing
1. Use the 1-d template [ -1,0,+1] to calculate the gradient in the GX and GY directions
2. Gradient amplitude values are used
3. The gradient map can be represented as a 8bit slice map, each of which corresponds to 0 or 1 (the author thinks the wheel
There are regions with strong gradients, so the first 4 bits of the gradient amplitude are represented by each point.
4. The gradient feature (8x8) is then extracted for each slice graph, which will eventually extract four facet features and merge
Is the Bing feature of the region.
PS. For the 4th step, the author uses an accelerated process here, that is, each time the next feature is computed, use
The value of the previous feature, you only need to do a [shift] and [or] operation.
If you move the red box up one line (that is, it does not contain a green box),
When the value of the vector is set to, then the eigenvalues of the red box position can be
Expressed as, at this point, the calculation feature is not duplicated
Cycle of the.
The training process (this Part I have not fully understood: some people understand the words remember to tell me, ah, thank you, I will continue to look down)
1. First, extract the Bing features of the positive and negative samples, input into the LINEAR-SVM to train a classifier, the classifier is normalized , as the first level of Cascade classifier.
2. Then, use this classifier to search for the training sample (larger image-guess may use image indentation detection), this time can get a lot of target box, using NMS
Suppress it, then select a smaller box as the second-level training sample, use this classifier to search for training negative cases (large image) search difficult samples (bootstrap strategy),
Generate a negative example of a second-level training sample.
3. Then, the positive and negative examples generated by the 2nd step are entered into the LINEAR-SVM to train a classifier as the second level of the Cascade classifier.
Classifier Normalization method
is a constant weight (linear classifier model), is a normalized dimension (to turn w into a few dimensions,
The author takes a value of 2), this theory is to do projections, also wood has to see proof
[Efficient] online structured output learning for keypoint-based object tracking.
continued-by [email protected] Dddz WDH 2015-01-19
Bing:binarized normed gradients features for target detection < reading notes 1>