In order for a dataset to be made from a video there must be at least one labeled frame in the video. However the check for a labeled frame is flawed. Although it is not possible to change frame (if the user does not subvert the JS checks) without entering a label for a model the frame is still marked as labeled. This means it is possible to have a frame, like the one shown below, that is seen as labeled. This means that it is now possible to create a dataset from this video. However this video does not actually have a single labeled frame.
In addition from what I can tell the only thing preventing the user from going to the next frame and entering more "empty bounding boxes" is the client side JS code. Perhaps there should be a check on the server side. I think such a check should only let one frame have incomplete boxes at a time. When there is one frame with incomplete bounding boxes the rest should have the creation of new bounding boxes disabled.
Consequences of Bug
As was said before this bug makes it possible to make an empty dataset (seen below). From this it is possible to train a model. In the interest of not wasting training time I have not tested training a model from this empty dataset. However I can only imagine that it would cause many errors.
Replication of Bug
Upload video
Open video labeler after processing complete
Create bounding box of any size and location
Return to home screen without entering label for bounding box
Note that video that has just been made now has 1 labeled frame and that it is now possible to make a dataset from said video
Bug
In order for a dataset to be made from a video there must be at least one labeled frame in the video. However the check for a labeled frame is flawed. Although it is not possible to change frame (if the user does not subvert the JS checks) without entering a label for a model the frame is still marked as labeled. This means it is possible to have a frame, like the one shown below, that is seen as labeled. This means that it is now possible to create a dataset from this video. However this video does not actually have a single labeled frame. In addition from what I can tell the only thing preventing the user from going to the next frame and entering more "empty bounding boxes" is the client side JS code. Perhaps there should be a check on the server side. I think such a check should only let one frame have incomplete boxes at a time. When there is one frame with incomplete bounding boxes the rest should have the creation of new bounding boxes disabled.
Consequences of Bug
As was said before this bug makes it possible to make an empty dataset (seen below). From this it is possible to train a model. In the interest of not wasting training time I have not tested training a model from this empty dataset. However I can only imagine that it would cause many errors.
Replication of Bug