large rotation steps do not add up to 360 deg.

[jwstolk]

It currently is not possible to return to horizontal or vertical using the large Z and V, or even the small X and C steps. This is caused by the large steps being 0.1 radians and the small steps 0.1 radians, which will never add up to 2*Pi.

proposed to use math.radians() in roLabelImg-master/libs/canvas.py as below:

            self.selectedShape.isRotated and not self.rotateOutOfBound(math.radians(45)):
            self.selectedShape.rotate(math.radians(45))
            self.shapeMoved.emit() 
            self.update()  
        elif key == Qt.Key_X and self.selectedShape and\
             self.selectedShape.isRotated and not self.rotateOutOfBound(math.radians(2.5)):
            self.selectedShape.rotate(math.radians(2.5)) 
            self.shapeMoved.emit()
            self.update()  
        elif key == Qt.Key_C and self.selectedShape and\
             self.selectedShape.isRotated and not self.rotateOutOfBound(-math.radians(2.5)):
            self.selectedShape.rotate(-math.radians(2.5)) 
            self.shapeMoved.emit()
            self.update()  
        elif key == Qt.Key_V and self.selectedShape and\
             self.selectedShape.isRotated and not self.rotateOutOfBound(-math.radians(45)):
            self.selectedShape.rotate(-math.radians(45))
            self.shapeMoved.emit()
            self.update()

This does 45deg large steps and 2.5 deg small steps, but anything reasonable will do.
It's also much easier to understand for someone to change the values.

[jwstolk]

(for info) My original plan was to label each object twice, once with a rectangle and once with a 45 degree rotated rectangle, make a 45-degree rotated version of the training images and train YOLO with both version of each image, with the appropriate bounding rectangle.

Then run each frame twice through YOLO, once normally and once rotated 45 deg. Then merge the resulting bounding boxes into an octagonal outline. This would need minimal changes to YOLO but would half the framerate and require each frame to be rotated 45 deg.

(Now thinking about a better solution that will require some larger changes to roLabelimg and YOLO.)

[jaychen1996-byte]

@jwstolk Hi,bro,How is your 45-degree rotating YOLO model trained?