Vocabulary optimization

[akunin]

Hi Mateu,
As incremental vocabulary for BOW takes a lot of memory for big maps - we are looking for a way to optimize it. Unfortunately the fixed one doesn't provide the same accuracy.
Currently new words are added for each processed image.
Whether do you have any parameter allowing us to generate new words 1/10 rate ? (I dont want to decrease linear update since we want to have all nodes/signature but reduce vocabulary size only).

Any other ideas to optimize it?
thanks

[matlabbe]

You can play with parameter Kp/NndrRatio (default 0.8) to adjust how many new words are created for each image. Increasing it will result in a smaller dictionary though loop closure detection recall may decrease. See also Figure 11 of that paper.

Decreasing the number of features used for dictionary (Kp/MaxFeatures, default 500) can be also another option.