rtabmap-console and TimeThr

[m.omar82]

Dear matlabbe
I am working rtabmap-console "from the last update" cons try to test SIFT with Lib6Outdoor even images, from the information display I get:
 iteration(363) loop(133) hyp(0.93) time=3.864293s/3.864617s *
 iteration(364) loop(134) hyp(0.93) time=4.184517s/4.184946s *
 iteration(365) loop(136) hyp(0.90) time=4.038019s/4.038360s

is it normal to get 3.8s or 4.1s and the -Rtabmap/TimeThr is 1400?
I use "-Kp/NNStrategy 3" for Brute Force matching, I know it takes a lot of time but is it rtab-map will decrease WM size to make the process time less than 1.4s?

[matlabbe]

Hi,

The problem with using BruteForce matching is when some signatures are retrieved from LTM, the retrieved words should be matched to dictionary again, thus with brute force matching we have almost double time required for an iteration (when signatures are retrieved). Looking the logs, the WM is effectively decreasing, but as we can see in the time plot when signatures are retrieved, we have double time required. When using KD-Trees, the overhead of matching retrieved words is much lower than with BruteForce, then real-time constraints can be satisfied.

Time:


WM:

[m.omar82]

Thanks for your replay,
1. that is mean BoW "Bag of Word" just represent the locations in WM?

2. How can make only one memory , just WM and no LTM?

3. from the Code "in Rtabmap.cpp", what functions I have to call to retrieve location from LTM to WM, for example I like to ignore the Bayesian filter result and like to select random location to be loop  closure for the current image, if am luck the random location already available in WM but if the random location is in LTM?

[matlabbe]

1. yes

2. set Rtabmap/TimeThr=0 and Rtabmap/MemoryThr=0, the vocabulary and WM will then increase indefinitely.

3. You can add them to reactivateIds list, so they will be retrieved (if not in WM) from LTM when reactivateSignatures() is called. The first ones in the list are loaded first, and limited to "Rtabmap/MaxRetrieved" (default 2).

cheers

[m.omar82]

THANK for your fast support