hslol / rotation-retinanet-pytorch

A Rotation RetinaNet Pytorch Implementation on HRSC and SSDD Dataset.

Hi, I'm Zhonghao Fang. 👋

• 😄 I’m currently working on Horizontal and Oriented Object Detection Based on Remote Sensing Images.
• 🔭 I’m currently focusing on BEV Perceptron and Occupancy Netowrk in Auto Driving.
• 📫 How to reach me: [email protected].
• ✨ HomePage: https://hslol.github.io/

rotation-retinanet-pytorch's Issues

ssdd 精度问题

博主你好，我在mmrotate上实验ssdd，发现map只能到80，而我看你的实验可以到84以上，想问一下你的实验和mmrotate里面的retina有什么区别吗

make 文件问题

博主你好，请问执行完make指令后，运行训练脚本仍然提示ModuleNotFoundError: No module named 'utils.rotation_overlaps.rbox_overlaps'该如何解决呢。

cpu运行问题

你好博主，可以在cpu上训练吗，如何调整，如果可以的话能否增添更新代码

ground_truth文件问题

你好，博主，可以提供以下txt格式的图片标注文件吗

horizontal_overlaps计算是否存在问题？

您好，看了您的代码，非常适合初学者读。这里有一个问题请教，网络产生的anchor是 [xc, yc, w, h, angle]的形式，由于gt也是 [xc, yc, w, h, angle]的形式，所以在计算horizontal_overlaps 的时候您将gt通过obb2hbb_oc转成了hbb的形式，但是anchor并没有转啊，这怎么求horizontal_overlaps的呢？

horizontal_overlaps = bbox_overlaps(
anchor.clone(), # generate anchor data copy
obb2hbb_oc(bbox_annotation[:, :-1]))

我理解应该是
horizontal_overlaps = bbox_overlaps(
obb2hbb_oc(anchor.clone()), # generate anchor data copy
obb2hbb_oc(bbox_annotation[:, :-1]))
请问我的理解是否存在偏差？

对poly2obb_np_oc注释的疑问

您好,感谢您的代码.我在自己的机器上跑通了您的代码,并且效果很好.学习您的代码过程中,我有一个小问题.
您的代码中poly2obb_np_oc是这样写的

def poly2obb_np_oc(poly):
    """ Modified !!
    Convert polygons to oriented bounding boxes.
    Args:
        polys (ndarray): [x0,y0,x1,y1,x2,y2,x3,y3]
    Returns:
        obbs (ndarray): [x_ctr,y_ctr,w,h,angle] modified -> [x_ctr, y_ctr, h, w, angle(radian)]
    """
    bboxps = np.array(poly).reshape((4, 2))
    rbbox = cv2.minAreaRect(bboxps)
    x, y, h, w, a = rbbox[0][0], rbbox[0][1], rbbox[1][0], rbbox[1][1], rbbox[2]
    # assert 0 < a <= 90, f'error from poly2obb_np_oc function.'
    if w < 2 or h < 2:
        return
    while not 0 < a <= 90:
        if a == -90:
            a += 180
        else:
            a += 90
            w, h = h, w
    a = a / 180 * np.pi
    assert 0 < a <= np.pi / 2
    return x, y, h, w, a

在cv2.__version__为4.5.5的环境中,我进行了测试.我认为
x, y, h, w, a = rbbox[0][0], rbbox[0][1], rbbox[1][0], rbbox[1][1], rbbox[2]
应该为
x, y, w, h, a = rbbox[0][0], rbbox[0][1], rbbox[1][0], rbbox[1][1], rbbox[2]
即w应该为rbbox[1][0],h应该为rbbox[1][1].并且,我在mmrotate上找到同样的函数

def poly2obb_np_oc(poly):
    """Convert polygons to oriented bounding boxes.
    Args:
        polys (ndarray): [x0,y0,x1,y1,x2,y2,x3,y3]
    Returns:
        obbs (ndarray): [x_ctr,y_ctr,w,h,angle]
    """
    bboxps = np.array(poly).reshape((4, 2))
    rbbox = cv2.minAreaRect(bboxps)
    x, y, w, h, a = rbbox[0][0], rbbox[0][1], rbbox[1][0], rbbox[1][1], rbbox[
        2]
    if w < 2 or h < 2:
        return
    while not 0 < a <= 90:
        if a == -90:
            a += 180
        else:
            a += 90
            w, h = h, w
    a = a / 180 * np.pi
    assert 0 < a <= np.pi / 2
    return x, y, w, h, a

mmrotate该函数实现
所以我很困惑您注释中的
obbs (ndarray): [x_ctr,y_ctr,w,h,angle] modified -> [x_ctr, y_ctr, h, w, angle(radian)].

How to predict multi-scale outputs?

Really appreciate your work. In your code, RetinaNet only predict a single scale bbox, so how to design a multi-scale output (like YOLO series)? I am actually concerned about how to code about loss and decoder in models/model.py.