1.yuv+gradient mag+gradient hist
2.用一阶通道特征就够了,高阶意义不大
3.对输入图像做线性或非线性变换后,使得诸如局部求和,直方图,haar-like(二阶)及其变种特征可以快速由积分图算出4.LUV转换的瓶颈在于开立方根的运算,如何近似?
5.梯度直方图计算trick:两个2维向量角度的近似可以由:在向量空间沿着单位圆寻找二向量的点积的最大值.(二分查找快速实现).
6.一般而言:(320*240)LUV 135fps,gradient magnitude 140 fps, gradient histogram (6 bins) 60 fps. --> all 10 channel: 40 fps. or 30 fps if pre-smoothing with r <=8;(exclude classifier)
7.soft cascade:a threshold is used after evaluation of every weak classifier.boost,特征的值需要被量化并且在开始训练时被cached.
8.scale:pre-smoothing with a radius 1 binomial filter(sigma = 0.87) improved performance.
9.classifier: more weak classifier,depth-2 trees gave the best performance rather than stumps.(choice of boosting algo does not matter)
10.displayed channels are slightly smaller than input image,不要用边界
11.hog or any other types of feature can be converted into channel features.