数据标注格式
Lmoer opened this issue · 6 comments
请问一下LZ，旋转框的数据标注格式是怎么样的，如果可以，对于完整的pipline，可否写一个类似于https://github.com/ultralytics/yolov3的pipline readme
LZ如果可以，对于完整的pipline，可否写一个类似于https://github.com/ultralytics/yolov3 的 pipline readme,代码看得我一头雾水
Dear @ming71 ,
Thank you for sharing your work.
I am not clear about the rbbox annotation of the HRSC dataset. I guess it is in [xc, yc, w, h, ang], but I do not know how w, h are measured, parallel to the image coordinates or from the cv2.minAreaRect, as well as the ang? Here is my dataset label format rbbox = [xc, yc, w, h, ang]:
rect = cv2.minAreaRect(hull)
xc, yc = rect[0][0], rect[0][1]
w, h = rect[1][0], rect[1][1]
ang = rect[2]
rbbox = [xc, yc, w, h, ang]
However, according to this source, the ang varies between -0 to -90 and the vertex arrangement is quite specific.
When you train the HRSC dataset:
-
wis always larger thanh? -
Where is the origin to measure the
ang(-pi/2, pi/2)?
Thank you for your time.
e is my
hello, u r right, the implementation here does not consider the periodicity of the angle and strictly defination the width and height, which may lead to ambiguity in the prediction results.I have not been able to optimize before.
As for the annotation format, I use the same format as yolov3, and the additional angle variable is introduced, specifically, (cx, cy, w, h, a), in which cx, cy, w, h are normalized to (0,1). More details can be found here.
e is my
hello, u r right, the implementation here does not consider the periodicity of the angle and strictly defination the width and height, which may lead to ambiguity in the prediction results.I have not been able to optimize before.
As for the annotation format, I use the same format as yolov3, and the additional angle variable is introduced, specifically, (cx, cy, w, h, a), in which cx, cy, w, h are normalized to (0,1). More details can be found here.
cx cy我知道，cx/1920 ，cy/1080，但是w和h怎么归一化呢，anchor是有角度的，总不能w/1920，h/1080吧
w和h也用图像尺度归一化,角度不用处理.