Spatial Pyramid Pooling in Deep Convolutional --- Spp_net

微软亚研院2015的一篇文章，优点是能够满足任意大小图像的输入。

主要思想：

（1）Spatial Pyramid Pooling Layer. 正是因为该层，才让Spp_net能够实现任意图片的输入，并且得到固定长度的特征向量：

stride和window的计算：

（2）Mapping a Window to Feature Maps. 将原图输入Spp_net后，通过下面图片中介绍的方法，能够将原图中的点映射到feature map上，为object detection打下基础：

主要代码实现（基于theano/keras）：

（1）spp_layer:

class SppLayer(Layer):    def __init__(self,bins,feature_map_size=0):        super(SppLayer,self).__init__()        self.strides = []        self.windows = []        self.a = feature_map_size#feature_map size        self.bins = bins        self.num_bins = len(bins)    def get_output(self,train):        self.input = self.get_input(train)        for i in range(self.num_bins):            self.strides.append(int(math.floor(self.a/self.bins[i])))            self.windows.append(int(math.ceil(self.a/self.bins[i])))        self.pooled_out = []        for j in range(self.num_bins):            self.pooled_out.append(downsample.max_pool_2d(input=self.input,                                                              ds=(self.windows[j],self.windows[j]),                                                              st=(self.strides[j],self.strides[j]),                                                              ignore_border=False))        for k in range(self.num_bins):            self.pooled_out[k] = self.pooled_out[k].flatten(2)            """            print self.windows[k]            print self.strides[k]            print 'K: '+str(k)            """        # batch_size * image_size        self.output = T.concatenate([self.pooled_out[0],self.pooled_out[1],self.pooled_out[2]],axis=1)        return self.output

(2)Mapping a Window to Feature Maps:

def window_to_feature_map(window_point_x1,window_point_y1,window_point_x2,window_point_y2,                          window_size_x,window_size_y,map_size_x,map_size_y):    map_point_x1 = window_point_x1*math.ceil(map_size_x/window_size_x)-1    map_point_y1 = window_point_y1*math.ceil(map_size_y/window_size_y)-1    map_point_x2 = window_point_x2*math.ceil(map_size_x/window_size_x)-1    map_point_y2 = window_point_y2*math.ceil(map_size_y/window_size_y)-1    return map_point_x1,map_point_y1,map_point_x2,map_point_y2