GPU-matching

Files:

1. nn.py - main file to run. There are 3 algorithms implemented, the nearest-neighbor matching from query set of descriptors to the database (DB), both for CPU and GPU:

   1. brute-force;
   2. Manhattan-distance based;
   3. LSH-based.

   For running on GPU numba python module is needed (conda install numba)

2. plot.py -

3. plot_descr_on_image.py - plots match of features from two images using saved before files (with ending .res.csv.)

4. r2d2_torch - copy of the depository (https://github.com/naver/r2d2) for extracting R2D2 descriptors in Pytorch (needs it to be installed, or already created files with descriptors could be used )

5. r2d2_torch/imgs/msk[2].png.*.r2d2 - descriptors extracted by R2D2, used for test run

To run, simply:

python nn.py -t TYPE -m MODE --rep N

where

TYPE is from ("cpu", "gpu")

MODE is from ("brute", "manh", "hash")

N - number to repeat (for collecting statistics, default 1).

Other parameters include:

1. using all images mentioned in code or only one for DB construction - --add_dbs 1.
2. if needed to save coordinates of correspondences --coords_save 1.

For extracting K features run from r2d2_torch:

python extract.py --model models/r2d2_WASF_N8_big.pt --images imgs/ms.png --top-k K

Technical details:

1. If needed only one set of parameters (the best), with the additional argument --final 1 will be taken only one specified in the code set.
2. The results are saved as .csv files in subfolder results with names {type}_{mode}.csv and {type}_{mode}.final.csv if it was the final one (with corresp. argument). For further details see argument of running nn.py.
3. Maximal allowed number of descriptors in DB is hard-coded in functions find_dist_linked and find_dist_manh.