We used the auto-detect bounding box function using algorithm MedianFlow. When a label is off the screen it still shows the label. How do I make it so it doesn’t auto-detect off the screen or delete the incorrect labels?
It ends up having labels that are of blank space as the object was rotated. Should I pause the auto-label during rotation?
As you are tracking, the OpenCV tracking algorithm is attempting to use the previous frame’s detection data to make determinations on what to track for each subsequent frame. Often OpenCV gets it right, sometimes it gets it wrong, and sometimes it gets things hilariously wrong. I recommend reading up about the different tracking algorithms by clicking on the “What’s this?” link below the algorithm selector in ftc-ml.
In general, the MedianFlow tracking algorithm is best when the object in the frame is moving relatively slowly and is not changing significantly. This makes MedianFlow a GREAT algorithm when tracking mostly-fixed-orientation objects, like images on a Signal Sleeve. However, when tracking objects that are constantly changing orientation (e.g. a non-uniform body rotating in the frame) I have found that the CSRT algorithm is a little more forgiving even though its bounding boxes are often a hot mess.
When performing tracking, it is ABSOLUTELY a good idea to pause tracking as soon as the bounding boxes do not look correct. When you pause tracking, you can edit the frame (and only that frame) that the tracker stopped on. If the tracker messed up several frames already or you couldn’t pause fast enough, you can always stop tracking - stopping the tracking allows you to go back and fix any frames you want. You can then resume tracking whenever you want; to do this, move to where you’d like to resume (or even restart) tracking - ensuring that the frame has correct labels on it - and press the “Start Tracking” button. The tracker will then use the current frame’s labels to begin/resume tracking the object(s) in the next (and subsequent) frames.
Remember, you can always delete labels by pressing the “Trash Can” icon for the label in the label list that corresponds with the bounding box in the image frame, and you can create labels by just clicking and dragging a box in the image. You can also edit the label’s bounding box by either editing the X and Y values in the table or you can drag the label’s corners (the black dots) in the image.