Update LFW results with changes from #50.

.gitignore

@@ -13,6 +13,8 @@ images/examples-aligned
 
 evaluation/*/*.csv
 evaluation/*/*.pdf
+evaluation/attic/*/*.csv
+evaluation/attic/*/*.pdf
 
 demos/web/bower_components
 demos/web/unknown*.npy

docs/accuracy.md

@@ -5,8 +5,8 @@ on the standard
 [LFW](http://vis-www.cs.umass.edu/lfw/results.html)
 benchmark.
 We had to fallback to using the deep funneled versions for
-152 of 13233 images because dlib failed to detect a face or landmarks.
-We obtain a mean accuracy of 0.8483 ± 0.0172 with an AUC of 0.923.
+58 of 13233 images because dlib failed to detect a face or landmarks.
+We obtain a mean accuracy of 0.8138 ± 0.0149 with an AUC of 0.893.
 For comparison, training with Google-scale data results in an
 accuracy of .9963 ± 0.009.
 

evaluation/attic/lfw.nn4.v1.epoch-177.orig-alignment/accuracies.txt

@@ -0,0 +1,13 @@
+fold, threshold, accuracy
+0, 0.71, 0.86
+1, 0.71, 0.83
+2, 0.77, 0.82
+3, 0.70, 0.84
+4, 0.68, 0.82
+5, 0.71, 0.85
+6, 0.69, 0.85
+7, 0.71, 0.85
+8, 0.71, 0.86
+9, 0.67, 0.88
+
+avg, 0.8483 +/- 0.0172

evaluation/lfw.nn4.v1.epoch-177/accuracies.txt

@@ -1,13 +1,13 @@
 fold, threshold, accuracy
-0, 0.71, 0.86
-1, 0.71, 0.83
-2, 0.77, 0.82
-3, 0.70, 0.84
-4, 0.68, 0.82
-5, 0.71, 0.85
-6, 0.69, 0.85
-7, 0.71, 0.85
-8, 0.71, 0.86
-9, 0.67, 0.88
+0, 0.68, 0.82
+1, 0.68, 0.81
+2, 0.68, 0.79
+3, 0.68, 0.80
+4, 0.68, 0.81
+5, 0.68, 0.83
+6, 0.68, 0.83
+7, 0.68, 0.82
+8, 0.68, 0.81
+9, 0.60, 0.80
 
-avg, 0.8483 +/- 0.0172
+avg, 0.8138 +/- 0.0149