We provide provide 3 separate tasks, inspired by different practical aspects of local feature descriptors. In depth details can be found on [1].
| task name | description |
|---|---|
verification |
measures how well a descriptor separates positive from negative pairs of patches |
matching |
measures how well a descriptor matches two images |
retrieval |
measures how well a descriptor retrieves similar patches from a large collection |
Description of the task definition files
For the verification task, for each split X, we provide 3 files:
| file name | description |
|---|---|
verif_pos_split-X.csv |
contains the positive pairs |
verif_neg_inter_split-X.csv |
contains negative pairs sampled from between sequences |
verif_neg_intra_split-X.csv |
contains negative pairs sampled from withing sequences |
In utils/tasks/ we provide the task definition files for both positive and negative pairs, which are of the form
Negative pairs:
s1,t1,idx1,s2,t2,idx2
v_birdwoman,4,251,i_resort,3,204
v_strand,1,1625,i_santuario,3,482
i_salon,0,1100,i_ajuntament,3,544
...
Positive pairs:
s1,t1,idx1,s2,t2,idx2
v_birdwoman,4,251,v_birdwoman,3,251
v_strand,1,1625,v_strand,3,1625
i_salon,0,1100,i_salon,3,1100
i_zion,3,1766,i_zion,0,1766
...
Each pair is defined by the items of the header
s1,t1,idx1,s2,t2,idx2 where s represents the sequence, t the
image id from which the patch is extracted (0-5 since in HPatch there are 6 images per sequence), and idx the index of the patch inside the
sequence.
For example the, patch v_birdwoman,4,251 is 252th patch from the
v_birdwoman sequence, extracted from the 5th image and the
v_birdwoman,0,251 is the same, patch, but extracted from the
ref image.
Note that you can visualise individual patches using the hpatch_vis.py script provided.
For the retrieval task, we provide two files per split X:
| file name | description |
|---|---|
retr_queries_split-X.csv |
contains the queries |
retr_distractors_split-X-X.csv |
contains the pool with the distractors |
Note that the file contents are slightly different than the ones in
the verification task:
s,idx
v_man,841
v_birdwoman,1758
v_coffeehouse,219
i_salon,986
v_yuri,9
i_ajuntament,799
since in the retrieval task, only the patches from the ref image are
used, and thus, there is no need for the t index.
The matching task, does not make use of definition files, since due to its nature there is no need for repeatability enforcement.