我个人有个比较困惑的地方，关于anchor_size，希望有朋友可以解答一下

[clw5180]

对于FPN版本，作者推荐的anchor size为[2, 4, 8, 16, 32]，和maskrcnn_benchmark默认的anchor size [32, 64, 128, 256, 512]差别较大，我看了一下生成anchor部分的代码还是有一定区别：

原始RPN：

def generate_anchors(
    stride=16, sizes=(32, 64, 128, 256, 512), aspect_ratios=(0.5, 1, 2)
):
    """Generates a matrix of anchor boxes in (x1, y1, x2, y2) format. Anchors
    are centered on stride / 2, have (approximate) sqrt areas of the specified
    sizes, and aspect ratios as given.
    """
    return _generate_anchors(
        stride,
        np.array(sizes, dtype=np.float) / stride,
        np.array(aspect_ratios, dtype=np.float),
    )

而对于RRPN，这里sizes少除了一个stride：

def generate_anchors(
    stride=16, sizes=(8, 16, 32), aspect_ratios=(0.5, 1, 2), angle=(-30.0, 0.0, 30.0, 60.0, 90.0, 120.0)
):
    """Generates a matrix of anchor boxes in (x, y, w, h, theta) format. Anchors
    are centered on stride / 2, have (approximate) sqrt areas of the specified
    sizes, and aspect ratios as given.
    """
    return generate_Ranchors(
        stride,
        aspect_ratios,
        sizes,
        angle
    )

而在函数内部， 对于RPN：

def _generate_anchors(base_size, scales, aspect_ratios):
    """Generate anchor (reference) windows by enumerating aspect ratios X
    scales wrt a reference (0, 0, base_size - 1, base_size - 1) window.
    """
    anchor = np.array([1, 1, base_size, base_size], dtype=np.float) - 1
    anchors = _ratio_enum(anchor, aspect_ratios)
    anchors = np.vstack(
        [_scale_enum(anchors[i, :], scales) for i in range(anchors.shape[0])]
    )
    return torch.from_numpy(anchors)

RRPN：

def generate_Ranchors(base_size=16, ratios=[0.2, 0.5, 1.0],
                     scales=2 ** np.arange(3, 6), angle=[-30.0, 0.0, 30.0, 60.0, 90.0, 120.0]):
    ######################################
    # Parameter:
    ######################################
    # ratio: [0.5, 1, 2]
    # scales: 2^3, 2^4, 2^5
    # angle: [-45, 0, 45]
    ######################################
    # (project in original patch)
    ######################################

    # [x_ctr, y_ctr, height, width, theta] anti-clock-wise angle

    base_anchor = np.array([base_size / 2 - 1, base_size / 2 - 1, base_size, base_size, 0], dtype=np.float32)
    ratio_anchors = _ratio_enum(base_anchor, ratios)
    scale_anchors = np.vstack([_scale_enum(ratio_anchors[i, :], scales) for i in range(ratio_anchors.shape[0])])
    anchors = np.vstack([_angle_enum(scale_anchors[i, :], angle) for i in range(scale_anchors.shape[0])])

    # discard
    # anchors[:, 2:4] = anchors[:, 3:1:-1]

    return torch.from_numpy(anchors)

对于RPN的写法，generate_anchors相当于传入的是size / stride，即[8, 8, 8, 8, 8]，即每个feature map上的size，如果换算到原图就要乘以各自层的stride，得到[32, 64, 128, 256, 512] 而对于RRPN这种写法，generate_anchors相当于传入的是size本身，即[2, 4, 8, 16, 32]，换算到原图的实际anchor_size应该也要乘一个stride，变成[8, 32, 128, 512, 2048]。不知这样理解是否正确？

[wnynotfaker]

你好，我也发现了这个问题，跟您想的差不多。我对代码进行了修改，使得每个特征图上对应一个尺度的ａｎｃｈｏｒ，即ＲＰＮ的写法

[tiny-cold-hands]

你好，我也发现了这个问题，跟您想的差不多。我对代码进行了修改，使得每个特征图上对应一个尺度的ａｎｃｈｏｒ，即ＲＰＮ的写法

你好，请问是将倒数第二行改为size/stride吗？配置文件的anchor参数如何设置呢？ def generate_anchors( stride=16, sizes=(8, 16, 32), aspect_ratios=(0.5, 1, 2), angle=(-30.0, 0.0, 30.0, 60.0, 90.0, 120.0) ): """Generates a matrix of anchor boxes in (x, y, w, h, theta) format. Anchors are centered on stride / 2, have (approximate) sqrt areas of the specified sizes, and aspect ratios as given. """ return generate_Ranchors( stride, aspect_ratios, sizes, angle )