search

robotology-legacy / wysiwyd

What You Say Is What You Did
http://wysiwyd.upf.edu
23 stars 16 forks source link

[pasar] Pasar is overfilling the ABM #95

Closed Tobias-Fischer closed 8 years ago

Tobias-Fischer commented 8 years ago

Hi @gregoire-pointeau,

@maxime-petit and I have discovered that pasar sends tens of snapshots to the ABM every second. Especially appearance and disappearance happens very often. Can you please reduce this amount? Otherwise it will fill up our hard drives very quickly .. and make narrative stories a mess as well.

gregoire-pointeau commented 8 years ago

I never witnessed this behavior. Can you send me a snapshot of your ABM to check what are the information that go in the ABM please? Here, we have a snapshot only at appearance or disappearance of an object, which is not very often.

maxime-petit commented 8 years ago

Actually, the appearance/disappearance is especially when the head is moving (look at the table, look at you, look at the table) which happens quite often in the demo. Because the appearance/disapperance in such case is just that the icub loose track of the object for some seconds, they are not "disappearing" (i.e. they are not removed from the table).

This thing cannot be replace by a check directly to OPC if an object was present/absent before/after an action? Or making sure the appearance/disappearance is a proper object add/remove to the working space.

./Max

gregoire-pointeau commented 8 years ago

We discussed a lot this problem at INSERM during the IE. What you are talking is supposed to be the role of attentionSelector which should set the presence to 0 / 0.5 / 1 and pasar checks for absence (0). Are you moving the head using attentionSelector ?

maxime-petit commented 8 years ago

No because we agree to not use it for the demo :)

We will try to increase pasar option about threshold for appearance/disapperance. By default it is respectively 1 and 2 seconds which in my opinion is way to short (a looking/pointing action take couple of second at least, appearance/disappearance should be robust to the "noise" from it). We will let you know

/Max

gregoire-pointeau commented 8 years ago

Ok. However, it is strange that you have tens of input per second. Can you send me a snapshot please ?

maxime-petit commented 8 years ago

Yop!

It is not tens input/second but it definitely have a "flickering" nature as soon as the head is moving (look or home action), as well as some disappearance/appearance separated by a few ms (5651/5652, 5679/5680)

look_pasar2 look_pasar

gregoire-pointeau commented 8 years ago

Thanks. It seems that you have between 7s and 10s (at mean) between each apperance and disapearance. Is it for the same object? I mean I coded a bit fast the thresholds for appereance and dissapearance, but I think they work.

If we decide to increase too much the thresholds, pasar won't detect that we bring a new object if we want to do the scenario with different objects.

pattacini commented 8 years ago

@maxime-petit can you double check if the object is actually disappearing during gaze activity? The tracking system, if working well, should prevent this problem.

Tobias-Fischer commented 8 years ago

Hi @pattacini, We've now added some lookAtAgent() all over the place, where the objects definitely disappear. It might be the reason. Our robot is having some hardware issues right now, we'll test soon ;)

gregoire-pointeau commented 8 years ago

Hi @tobias, did you have time to test it?

Tobias-Fischer commented 8 years ago

Hi @gregoire-pointeau, Yes it seems to be okay. Thanks :)