search

TWOEARS / blackboard-system

Two!Ears Auditory Model - Blackboard system module
http://docs.twoears.eu/en/latest/blackboard/
GNU General Public License v2.0
3 stars 2 forks source link

Work with short signals #15

Open hagenw opened 8 years ago

hagenw commented 8 years ago

This is again on getting the correct signals in the Blackboard.

Note: you have to update the internal database before executing the following things.

Please checkout the ring1 branch in TWOEARS/examples, got to the folder casa_kopco and run

>> localise;

------------------------------------------------------------------------------------
Reference target angle:   0 degrees
------------------------------------------------------------------------------------
Localised source angle:
BlockTime   PerceivedAzimuth    (head orient., relative azimuth)    Probability
------------------------------------------------------------------------------------
------------------------------------------------------------------------------------
Mean localisation error: NaN
------------------------------------------------------------------------------------

We set the length of the Binaural simulator to 1s and the localisation model works on 0.5s long signals, so it should be able to work, but it isn't.

Now checkout the ring2 branch, where we set the simulation length to 1.1s and run it again:

>> localise;

------------------------------------------------------------------------------------
Reference target angle:   0 degrees
------------------------------------------------------------------------------------
Localised source angle:
BlockTime   PerceivedAzimuth    (head orient., relative azimuth)    Probability
------------------------------------------------------------------------------------
  0.56       10 degrees     (  0 degrees,     10 degrees)   0.33
------------------------------------------------------------------------------------
Mean localisation error: 10
------------------------------------------------------------------------------------

Then it is working.

But the problem is that setting the time to a longer value does not always help. In the case of the real Kopco scenario we don't have a ring buffer (looping of the signal), but it is played just once. So, if we leave the time at 1.1s, but change the buffer to fifo the following will happen. Checkout the kopco branch and run

>> localise;

------------------------------------------------------------------------------------
Reference target angle:   0 degrees
------------------------------------------------------------------------------------
Localised source angle:
BlockTime   PerceivedAzimuth    (head orient., relative azimuth)    Probability
------------------------------------------------------------------------------------
------------------------------------------------------------------------------------
Mean localisation error: NaN
------------------------------------------------------------------------------------

As I have no clue how the signal and block based processing is handled, it would be great if someone who implemented it could have a look at this and make a proposal how to solve this problem.

/cc @Hardcorehobel @fietew @ningma97

hagenw commented 8 years ago

This was also affected by my usage of an outdated version of TWOEARS/blackboard-system. The differences between the branches ring1 and ring2 is now gone and both of them work.

However, the problem with switching to fifo in the kopco branch remains.