Why is it defined as a "Segmentation" game?

Isn't getting AoP the ultimate goal? Why should the head of the fetus simulate an ellipse? Is it not reasonable to directly calculate the tangent? Can classification be used?

Yes, the ultimate goal is to calculate AoP, and the ellipsoid-like shape of the fetal region serves as prior knowledge to facilitate the calculation of AoP. Of course, researchers can directly calculate the tangent as well. Classification algorithms can also be employed.

The ultimate goal is to achieve AOP. Usually this kind of task is to first segment the target and then estimate the parameters. According to the definition of AOP and the characteristics of the segmentation object, we plan to segment the FH-PS first, then fit the segmented area with an ellipse, and finally calculate the relevant parameters according to the endpoint of the major axis of the ellipse.

I'm actually questioning your scoring method. If the goal is AoP, first of all, your training set does not provide data for reference; second, your proposed angle calculation method is vague. Assuming that the segmentation is exactly the same as the mask you provided, will the angles calculated by each contestant be the same? How do you define the best AoP? Is the competition the segmentation accuracy or the calculation method?

Hi Xuchu,

Thanks for your questions.

The AOP calculation process uses three image layers, of which the first layer is the background layer, the second layer is the pubic joint, and the third layer is the fetal head layer. Step 1: Find all ellipses of the pubic joint and fetal head layers separately using cv2.findContours. Step 2: Traverse the center coordinates, major axis, minor axis, and rotation angle of the largest ellipse area in each layer respectively using cv2.fitEllipse. Step 3: Find the major axis of the pubic joint. Step 4: Find the tangential point of the fetal head. Step 5: Calculate the angle. If the segmentation is exactly the same as the mask we provided, the angles calculated by each contestant is the same.

So the competition considers both of the segmentation accuracy and the calculation method.

If you have further questions, let me know.

Jieyun Bai

No idea how to calculate the tangent passing through the major axis. I don't think it would be a big issue if you released only the evaluation code—at least the detailed algorithm to calculate it. For example, we have no idea if the background is used to calculate the Haussdroff distance and dice. What type of distance is used for the Haussdroff distance? Too much confusion!!!