Histogram of Oriented Gradients

1. Some Challenges

• Find robust feature set that allows object form to be discriminated.

• Challenges:

  • Wide range of pose and large variations in appearances 我们想得到比较有判别力的特征，他的形态会发生比较大的变化
  • Cluttered backgrounds under different illumination 背景车的变化，风吹树的变化，光照的变化
  • “Speed” for mobile vision 希望实时检测，对速度有很高的要求

• Local object appearance and shape can often be characterized rather well by the distribution of local intensity gradients or edge directions.局部梯度增强或者边缘检测

2. Histogram of Oriented Gradients

• Dividing the image window into small spatial regions (cells)
• Cells can be either rectangle or radial（径向的）.
• Each cell accumulating a weighted local 1-D histogram of gradient directions over the pixels of the cell. 计算每个像素的梯度，将梯度的方向量化为K个，然后将每个单元内相同的梯度方向的梯度幅值相加得到该方向的梯度强度，也就是直方图每个bin的值。

3. Normalization

• For better invariance to illumination and shadowing. it is useful to contrast-normalize the local responses before using them.

• Accumulate local histogram “energy” over a larger regions (“blocks”) to normalize all of the cells in the block.

4. Visualizing HoG

5. Difference between HoG and SIFT

• HoG is usually used to describe entire images. 即一下子计算了整张图的梯度
  SIFT is used for key point matching

• SIFT histograms are oriented 朝向towards the dominant gradient主方向.（提高了对旋转的鲁棒性）
  HoG is not.

• SIFT descriptors use varying scales to compute multiple descriptors. 提高对尺度鲁棒性

• HoGgradients are normalized using neighborhood bins（block内的直方图之和）.