Tarapio
Tarapio is a tool for tie points computation. Should be a faster and somehow extended alternative to the original tool developed for MICMAC, tapioca.
It gets more speed in this way:
- using OpenCV as backend for keypoints and descriptors computation
- using Flann for fast nearest neighbor finding in the feature space. Flann also allows for some level of approximation, permitting to achieve better performance (if an approximation can be accounted)
- using a disk cache for saving temporary files that can be reused at further tarapio runs.
- avoiding to compute two times the same match (if image A is matched with B then the inverse is not repeated)
- If image A have 25245 keypoints while B have 1245 keypoints the image A will be used for constructing the Flann kd-tree and the keypoints of B will be used as query points. And NOT vice-versa. In fact building the index is a fast operation while querying is not so fast.
Dependencies
A couple of dependencies:
- OpenCV
- Flann (probably if you have opencv you yet have Flann installed)
- Boost (used for parsing the regex at input)
- OpenMP that is used for parallelizing the computations
- A recent compiler. Actually some C++2011 standards have been used (mainly
shared_ptr- the boost version of it could be used actually). I wrote it using gcc-4.7.3, but gcc-4.6.* can compile it
Syntax
tarapio masterimgs [options]
where masterimgs is a regex pattern for matching the master images
tarapio permits to use a -q option for specifying a secondary set of query images. If master images are A and B, while query images are C, D and E, it will compute these matches:
A <-> C
A <-> D
A <-> E
B <-> C
B <-> D
B <-> E
See tarapio -h for further options
Notes
- The homologous points are saved in the Homol directory, as expected by Tapas. But remember to call Tapas with the ExpTxt=1 option! Because the points are saved as ascii files.
- The multiscale approach must still be implemented.
- The caching system is far away from being perfect.