Application of non-maximal value suppression in object detection

application of non-maximal value suppression in object detection

The application of NMS algorithm in object detection is introduced according to the source code of py_cpu_nms.py faster-rcnn. After the RPN layer in the FASTER-RCNN, some boundingbox and boundingbox corresponding to a certain class of scores (confidence degree) are obtained. So you can remove the box with a larger overlap value based on the NMS.

Import NumPy as NP def PY_CPU_NMS (Dets, Thresh): "" "Pure Python NMS baseline." "
    #dets: N*m,n is the number of bbox, the first 4 bits of M are corresponding (X1,y1,x2,y2), the 5th bit is the corresponding score #thresh: 0.3,0.5 .... X1 = dets[:, 0] y1 = dets[:, 1] x2 = dets[:, 2] y2 = dets[:, 3] scores = dets[:, 4] areas = (x2-x1 + 1) * (y2-y1 + 1) #求每个bbox的面积 order = Scores.argsort () [:: -1] #对分数进行倒排序 keep = [] #用来保存最后留下来的bbox while order.
        Size > 0:i = order[0] #无条件保留每次迭代中置信度最高的bbox keep.append (i) #计算置信度最高的bbox和其他剩下bbox之间的交叉区域 xx1 = Np.maximum (X1[i], x1[order[1:]]) yy1 = Np.maximum (Y1[i], y1[order[1:]) xx2 = Np.minimum (X2[i], x2
        [order[1:]])
        Yy2 = Np.minimum (Y2[i], y2[order[1:]) #计算置信度高的bbox和其他剩下bbox之间交叉区域的面积 w = np.maximum (0.0, xx2-xx1 + 1)  h = np.maximum (0.0, yy2-yy1 + 1) Inter = w * H #求交叉区域的面积占两者 (High confidence bbox and other bbox) area and the mandatory OVR = Inter/(Areas[i] + areas[order[1:]]-Inter) #保留ovr小于thresh的bbox, inInto the next iteration. Inds = Np.where (OVR <= thresh) [0] #因为ovr中的索引不包括order [0] so move backward one order = order[inds + 1] return ke
 Ep

The following explains the principle of using NMS to filter areas. When a window area is considered most likely to represent an object, then a proposal with a large cross area of the window area can be considered not to be the desired window area.

The above is a personal understanding of the NMS in object detection, if there is something wrong, I would like to point out. Another is to give a person feel that the NMS explained a better connection. Http://www.cnblogs.com/liekkas0626/p/5219244.html