关于siamrpn的一些想法

先看看label

def generate_labels(all_anchors, target_bbox):
    """
    There is no negative pairs in SiamRPN
    Args:
        anchors: all anchors scattered in detection frame
                 from `scatter_anchors` method
        bbox: ground truth bbox utils in detection
              frame, xyxy format

    Returns:
        cls: (K, 17, 17) cls branch label
        delta: (4, K, 17, 17) reg branch label
        delta_weight: (K, 17, 17) reg loss normalization factor
    """
    K, ft_size = all_anchors[0].shape[1], all_anchors[0].shape[2]
    tcx, tcy, tw, th = corner2center(target_bbox)

    # -1 ignore 0 negative 1 positive
    cls = -1 * np.ones((K, ft_size, ft_size), dtype=np.int64)
    delta = np.zeros((4, K, ft_size, ft_size), dtype=np.float32)
    delta_weight = np.zeros((K, ft_size, ft_size), dtype=np.float32)

    anchor_corner, anchor_center = all_anchors[0], all_anchors[1]

    x1, y1, x2, y2 = anchor_corner[0], anchor_corner[1], \
                     anchor_corner[2], anchor_corner[3]

    cx, cy, w, h   = anchor_center[0], anchor_center[1], \
                     anchor_center[2], anchor_center[3]
    # delta: (4, K, 17, 17) broadcast
    delta[0] = (tcx - cx) / w
    delta[1] = (tcy - cy) / h
    delta[2] = np.log(tw / w)
    delta[3] = np.log(th / h)
    # calculate IoU between target_bbox and all anchors
    overlap = IoU([x1, y1, x2, y2], target_bbox)

    pos = np.where(overlap > cfg.TRAIN.THR_HIGH)
    neg = np.where(overlap < cfg.TRAIN.THR_LOW)

    def select(position, keep_num=16):
        num = position[0].shape[0]
        if num <= keep_num:
            return position, num
        slt = np.arange(num)
        np.random.shuffle(slt)
        slt = slt[:keep_num]
        return tuple(p[slt] for p in position), keep_num

    # make sure at most 16 positive samples and totally 64 sam-
    # ples from one training pair as paper said
    pos, pos_num = select(pos, cfg.TRAIN.POS_NUM)
    neg, neg_num = select(neg, cfg.TRAIN.TOTAL_NUM - cfg.TRAIN.POS_NUM)

    # print("pos_num: ", pos_num)
    cls[pos] = 1
    cls[neg] = 0
    delta_weight[pos] = 1. / (pos_num + 1e-6)  # avoid ZeroDivisionError
    # (K,17,17) (4,K,17,17) (K,17,17) (1,)
    return cls, delta, delta_weight

当只有一张search时，
对于regression部分的label而言，形状为(4, K, 17, 17)，在这段代码里就是delta，其实就是对(K * 17 * 17)这么多个锚框统一都用论文中的那个公式计算，这样理论上无论选中哪个框都能正确地平移到物体那里，

delta_weight形状是(K,17,17)，作用是屏蔽负样本和模糊样本，只取正样本的regression进行求loss，实现上就是正样本的点有对应非0系数，代码里是正样本数目的倒数，负样本和模糊样本对应点系数为0

regression的loss

下面是regression的求loss过程
先求全部锚框的regression的loss
然后沿着regression维的方向求和，这样得到的featureMap形状为(N,K,17,17)，内容就是每个点装着自身4个offset的loss的和
最后乘上mask，只计算正样本，也就是iou大的锚框的regression的loss

classification的loss

模糊样本的label为-1，正样本为1,负样本为0，但计算loss的时候只计算正样本和负样本，而且是分别求loss，权重各占50%，都用同种loss，