Replies: 6 comments 24 replies
-
先检测旋转框，然后对四边进行一次分类如果我想在 rotate-retinanet 上添加头部检测的功能，是不是只要在 retinanet 的头部新增一个头部的分类的编码就行？ |
Beta Was this translation helpful? Give feedback.
-
|
如果方便的话，非常欢迎您提PR来完善mmrotate这部分功能。 |
Beta Was this translation helpful? Give feedback.
-
您是否有可供参考的开源代码~ |
Beta Was this translation helpful? Give feedback.
-
请问您有更新 mAP 的计算方式不哇？ 360 度时您是怎么处理的？ |
Beta Was this translation helpful? Give feedback.
-
def norm_angle(angle, angle_range):
"""Limit the range of angles.
Args:
angle (ndarray): shape(n, ).
angle_range (Str): angle representations.
Returns:
angle (ndarray): shape(n, ).
"""
if angle_range == 'oc':
return angle
elif angle_range == 'le135':
return (angle + np.pi / 4) % np.pi - np.pi / 4
elif angle_range == 'le90':
return (angle + np.pi / 2) % np.pi - np.pi / 2
else:
print('Not yet implemented.') |
Beta Was this translation helpful? Give feedback.
-
|
mAP那個不確定您的意思是什麼，不過在我的task中我依然採用DOTAMetric的評估方式，具體設置如下圖 然後pipline的設置會像是這樣 |
Beta Was this translation helpful? Give feedback.
-
|
我将 norm_angle 改成 如下，新增 angle_range == 'h180' 的情况，感觉写的不对，没有深刻理解 @qwert31639 , 虽然训练时跑起来了~ 感觉norm_angle写错了，虽然训练起来了 def norm_angle(angle, angle_range):
"""Limit the range of angles.
Args:
angle (ndarray): shape(n, ).
angle_range (Str): angle representations.
Returns:
angle (ndarray): shape(n, ).
"""
if angle_range == 'oc':
return angle
elif angle_range == 'le135':
return (angle + np.pi / 4) % np.pi - np.pi / 4
elif angle_range == 'le90':
return (angle + np.pi / 2) % np.pi - np.pi / 2
elif angle_range == 'h180':
return (angle + np.pi) % np.pi - np.pi
else:
print('Not yet implemented.')使用了 projects/headet/structures/bbox/rotated_boxes.py#L24 mmrotate/structures/bbox/transforms.py#L22 python tools/train.py projects/headet/configs/rotated_retinanet/rotated-retinanet-rbox-h180_r50_fpn_1x_dota.py
|
Beta Was this translation helpful? Give feedback.
-
|
原本1.x的code採用了cv2.minAreaRect，這會讓我找不到原始第一條邊的位置(也許大神有其他的解法)，所以我從新定義flip、rotate等函數，規則就是按照+-180度的邏輯走的，自然後面輸出的角度就不需要正規畫的動作 |
Beta Was this translation helpful? Give feedback.
-
|
h180 为什么是 return angle而不是 return (angle + np.pi) % (2*np.pi) - np.pi第二种方式的return 的范围才是 -pi,pi,第一种return角度的范围不是没有限制吗？ 可能我没有上下文 |
Beta Was this translation helpful? Give feedback.
-
感觉是不是下面这样也可以哇~ -90
|
|
+-180---------0
|
|
90
|
Beta Was this translation helpful? Give feedback.
-
|
的確，這個部份其實我沒有寫得很好，我就只是把正規化的邏輯放在了每一次旋轉或翻轉中比如說 |
Beta Was this translation helpful? Give feedback.
-
|
谢谢~ |
Beta Was this translation helpful? Give feedback.
-
|
角度预测出来之后，什么时候用到以下代码呢？ @register_box_converter(RotatedBoxes, QuadriBoxes, force=True)
def rbox2qbox(boxes: Tensor) -> Tensor:
"""copy from mmrotate/structures/bbox/box_converters.py Convert rotated
boxes to quadrilateral boxes.
Args:
boxes (Tensor): Rotated box tensor with shape of (..., 5).
Returns:
Tensor: Quadrilateral box tensor with shape of (..., 8).
"""
centerx, centery, w, h, theta = torch.split(boxes, (1, 1, 1, 1, 1), dim=-1)
# cos_value, sin_value = torch.cos(theta), torch.sin(theta)
cosa = torch.cos(theta)
sina = torch.sin(theta)
wx, wy = w / 2 * cosa, w / 2 * sina
hx, hy = -h / 2 * sina, h / 2 * cosa
p1x, p1y = centerx - wx - hx, centery - wy - hy
p2x, p2y = centerx + wx - hx, centery + wy - hy
p3x, p3y = centerx + wx + hx, centery + wy + hy
p4x, p4y = centerx - wx + hx, centery - wy + hy
return torch.stack([p1x, p1y, p2x, p2y, p3x, p3y, p4x, p4y], dim=-1) |
Beta Was this translation helpful? Give feedback.
-
|
其實那一段只是方便我看val的狀況而已，對於訓練沒有太大幫助，實際上也只有多畫了一個點 |
Beta Was this translation helpful? Give feedback.
-
|
thanks |
Beta Was this translation helpful? Give feedback.
-
qbox2rbox 的函数是不是也需要重写 |
Beta Was this translation helpful? Give feedback.
-
|
感觉需要修改一下 @register_box_converter(QuadriBoxes, RotatedBoxes)
def qbox2rbox(boxes: Tensor) -> Tensor:
"""Convert quadrilateral boxes to rotated boxes.
Args:
boxes (Tensor): Quadrilateral box tensor with shape of (..., 8).
Returns:
Tensor: Rotated box tensor with shape of (..., 5).
"""
# TODO support tensor-based minAreaRect later
original_shape = boxes.shape[:-1]
points = boxes.cpu().numpy().reshape(-1, 4, 2)
rboxes = []
for pts in points:
(x, y), (w, h), angle = cv2.minAreaRect(pts)
rboxes.append([x, y, w, h, angle / 180 * np.pi])
rboxes = boxes.new_tensor(rboxes)
return rboxes.view(*original_shape, 5) |
Beta Was this translation helpful? Give feedback.
-
|
@vansin 目前這個函數沒有使用到，但如果需要改可以參考 |
Beta Was this translation helpful? Give feedback.
-
|
剛剛稍微改了一下(還沒寫unit test)，大概是這樣 @register_box_converter(QuadriBoxes,RotatedBoxes, force=True)
def qbox2rbox(boxes: Tensor) -> Tensor:
"""Convert quadrilateral boxes to rotated boxes.
Args:
boxes (Tensor): Quadrilateral box tensor with shape of (..., 8).
Returns:
Tensor: Rotated box tensor with shape of (..., 5).
"""
points = boxes.cpu().numpy().reshape(-1, 4, 2)
cxs = points[...,0].mean()
cys = points[...,1].mean()
widthes= np.linalg.norm(points[:,1,:] - points[:,0,:])
heights= np.linalg.norm(points[:,2,:] - points[:,1,:])
thetas = np.arctan2((points[:,1,0]- points[:,2,0]),
(points[:,1,1]- points[:,2,1]))
thetas = np.where(thetas >= 0, np.pi - thetas, -(np.pi + thetas))
return torch.stack([cxs, cys, widthes, heights, thetas], dim=-1) |
Beta Was this translation helpful? Give feedback.
-
|
Is this still an open issue, or has anyone solved it? |
Beta Was this translation helpful? Give feedback.
-
|
请问这个问题解决了吗 |
Beta Was this translation helpful? Give feedback.
-
如何在 MMRotate 中支持 360°的旋转目标检测？
How to support 360° rotate target detection in MMRotate?
背景动机
现有的 MMRotate 的算法检测出来的旋转框，无论是 OpenCV表示法还是长边表示法，表示的范围只有180°，不能区分旋转框的头部和尾部。而头部和尾部信息在现实生活中是非常重要，例如区分出遥感图像中车辆的头部和尾部，飞机的头部和尾部时极具价值的。所以希望能够在 MMRotate 中实现 360°的旋转目标检测。
相关论文
(以下论文来自 yangxue 大佬的推荐)
（比较推荐）根据头部定义框，框的角度定义范围就是0-360，然后iou、nms以及不同框的转换函数对应改一下就行（iou应该直接用都不用改），参考文献：
EAGLE: Large-Scale Vehicle Detection Dataset in Real-World Scenarios using Aerial Imagery
DRBox-v2: An improved detector with rotatable boxes for target detection in SAR images
先检测旋转框，然后对四边进行一次分类，分出头部所在的边，参考文献：
On the Arbitrary-Oriented Object Detection: Classification based Approaches Revisited
https://github.com/yanqingyao1994/QPDet
数据集约定
在本部分，如何编码头部信息比较好？
DOTA 格式(起点顺时针定义法)
强制约定A为物体的起始点，A、B、C和D点顺时针排布。此时边AB物体的头部，在DOTA格式的数据集中就不必要标出物体的头部了。
评测指标
在经典的mAP的基础，类别，iou＞0.5，头部三个条件满足才算TP
旋转目标检测头部的可视化
此部分时实现旋转目标检测的头部的可视化~
Beta Was this translation helpful? Give feedback.
All reactions