Hey @ydayer, There are several annotations that are `NaN` Check a snippet that shows some samples that are Nan in the world set ```python >>> from bob.db.meds.database import Database >>> db = Database(protocol="verification_fold1", original_directory="") >>> samples = [[i, x.key,x.annotations] for i,x in enumerate(db.background_model_samples())] >>> print(samples[119]) [119, './data/ab/S200-01-t10_01.jpg', {'leye': (nan, nan), 'reye': (nan, nan)}] >>> # Prining all samples that has Nan annotations >>> print([[s[0],s[1]] for s in samples if numpy.isnan(s[2]['leye'][0])]) [[82, './data/ab/S131-01-t10_01.jpg'], [112, './data/ab/S180-01-t10_01.jpg'], [119, './data/ab/S200-01-t10_01.jpg'], [139, './data/ac/S237-01-t10_01.jpg'], [144, './data/ac/S245-01-t10_01.jpg'], [170, './data/ac/S290-01-t10_01.jpg'], [207, './data/ad/S362-01-t10_01.jpg']] ``` Do you have some time to look into that? This invalidates all the experiments we've done so far :-( Thanks