python 不使用for计算两组、多个矩形两两间的iou

解决问题： 不使用for计算两组、多个矩形两两间的iou

使用numpy广播的方法，在python程序中并不建议使用for语句，python中的for语句耗时较多，如果使用numpy广播的思想将会提速不少。

代码：

def calc_iou(bbox1, bbox2):
    if not isinstance(bbox1, np.ndarray):
        bbox1 = np.array(bbox1)
    if not isinstance(bbox2, np.ndarray):
        bbox2 = np.array(bbox2)
    xmin1, ymin1, xmax1, ymax1, = np.split(bbox1, 4, axis=-1)
    xmin2, ymin2, xmax2, ymax2, = np.split(bbox2, 4, axis=-1)

    area1 = (xmax1 - xmin1) * (ymax1 - ymin1)
    area2 = (xmax2 - xmin2) * (ymax2 - ymin2)

    ymin = np.maximum(ymin1, np.squeeze(ymin2, axis=-1))
    xmin = np.maximum(xmin1, np.squeeze(xmin2, axis=-1))
    ymax = np.minimum(ymax1, np.squeeze(ymax2, axis=-1))
    xmax = np.minimum(xmax1, np.squeeze(xmax2, axis=-1))

    h = np.maximum(ymax - ymin, 0)
    w = np.maximum(xmax - xmin, 0)
    intersect = h * w

    union = area1 + np.squeeze(area2, axis=-1) - intersect
    return intersect / union

 

程序中输入为多个矩形[xmin, ymin, xmax,ymax]格式的数组或者list，输出为numpy格式，例：输入的shape为（3， 4）、（5，4）则输出为（3， 5）各个位置为boxes间相互的iou值。后面会卡一个iou的阈值，然后就可以将满足条件的索引取出。如：

def delete_bbox(bbox1, bbox2, roi_bbox1, roi_bbox2, class1, class2, idx1, idx2, iou_value):
    idx = np.where(iou_value > 0.4)
    left_idx = idx[0]
    right_idx = idx[1]
    left = roi_bbox1[left_idx]
    right = roi_bbox2[right_idx]
    xmin1, ymin1, xmax1, ymax1, = np.split(left, 4, axis=-1)
    xmin2, ymin2, xmax2, ymax2, = np.split(right, 4, axis=-1)
    left_area = (xmax1 - xmin1) * (ymax1 - ymin1)
    right_area = (xmax2 - xmin2) * (ymax2 - ymin2)
    left_idx = left_idx[np.squeeze(left_area < right_area, axis=-1)]#小的被删
    right_idx = right_idx[np.squeeze(left_area > right_area, axis=-1)]

    bbox1 = np.delete(bbox1, idx1[left_idx], 0)
    class1 = np.delete(class1, idx1[left_idx])
    bbox2 = np.delete(bbox2, idx2[right_idx], 0)
    class2 = np.delete(class2, idx2[right_idx])

    return bbox1, bbox2, class1, class2

IOU计算原理：

ymin = np.maximum(ymin1, np.squeeze(ymin2, axis=-1))

xmin = np.maximum(xmin1, np.squeeze(xmin2, axis=-1))

ymax = np.minimum(ymax1, np.squeeze(ymax2, axis=-1))

xmax = np.minimum(xmax1, np.squeeze(xmax2, axis=-1))

h = np.maximum(ymax - ymin, 0)

w = np.maximum(xmax - xmin, 0)

intersect = h * w

计算矩形间min的最大值，max的最小值，如果ymax-ymin值大于0则如左图所示，如果小于0则如右图所示