MatchLab-Imperial / keras_triplet_descriptor

Baseline to denoise + learn descriptors in N-HPatches
17 stars 19 forks source link

Units in Verification #2

Closed martinferianc closed 5 years ago

martinferianc commented 5 years ago

Hi,

what are the units in the verification step, please?

E.g.:

>> Running HPatch evaluation for custom
>> Please wait, loading the descriptor files...
>> Descriptor files loaded.
>> Evaluating verification task
Processing verification task 1/3 : 100% 1000000/1000000 [01:29<00:00, 11152.96it/s]
Processing verification task 2/3 : 100% 1000000/1000000 [01:30<00:00, 11071.81it/s]
Processing verification task 3/3 : 100% 1000000/1000000 [01:31<00:00, 10930.76it/s]
>> Verification task finished in 280 secs  
Verification task results:
CUSTOM - Balanced variant (auc) 
Noise       Inter     Intra
-------  --------  --------
Easy     0.919498  0.887013
Hard     0.899671  0.8601
Tough    0.871768  0.825558
CUSTOM - Imbalanced variant (ap) 
Noise       Inter     Intra
-------  --------  --------
Easy     0.806251  0.721069
Hard     0.738682  0.627186
Tough    0.657986  0.536444
alopezgit commented 5 years ago

I will update the repo with explanation of the evaluation protocol followed soon. The results are the Area Under the Curve (AUC) of the ROC curve for the balanced case and mean Average Precision (Precision-Recall explanation) for the other case. It is explained in the HPatches paper along with the tasks,

martinferianc commented 5 years ago

Got it, thanks!