Phantom contours with MovementBasedSegmenter


#1

Hi guys

We did a test with the MBS by trying to classify a 3.5cm by 3.5cm colored square piece of paper. What we have noticed is the following:
If the watcher is added and no interaction is taking place, nothing happens (as intended).
Then, when the square is placed in the middle of the workspace, the MBS starts continuously registering contours, eventhough nothing changes. Once we placed the square, the MBS found 156 contours in 2min 46s without any further interaction.

For that, we use a MBS with the following params:
params: {
neural_network_name: ‘color-classifier’,
filter_rect: {
min_ratio: 0.9, // (0, 1], the closer to 1, the more “square” the contour must be
max_ratio: 1, // same as above
},
filter_thresh: 200
},
min and max ration has been chosen high to really only classify the colored squares, since they are square. The filter_thresh has also been chosen high to prevent unnessecary calls to the model.

Heres a screenshot of our setup from the /web/camera.html:
MBS-test-setup

Update: We have found the parameter filter_area: { min_ratio: int, max_ratio: int} which was missing in the api documentation. With this, we could restrict the amount of processed contours. But still, it seems as if the filter_rect and filter_thresh are not working as intended and the MBS is triggered unnecessarily often.

Update 2: It seems that the onLocation callback is also never invoked. On the simulator it is invoked, on the device it is not even though it sometimes invokes the onClassified callback on phantom contours. Also, the simulator passes the classification tag as string, while on the device it is a number.

Update 3: After adjusting the filter_area to only pick up our squares, we still get phantom contours, this time however always at position (-71.4506512963992:-53.20410927925579). I don’t know how it is possible to have a centerpoint of an object outside of the watcher area (negative coordinates). Has anyone ever noticed something similar?


#2

Hi @Yves

Could you please repost the image, we cannot open it.


#3

Hi @vsirbu

I’ve edited the post to reupload the image.