Hi, you really done a great job!
I re-implement your project with C++ code and got almost the same performance as yours in LFW.
But when I used it to extract features and test with cosine similarity in CASIA database, the accuracy is only about 80%.
The most of the wrong classification caused by large headpose and blurred faces.
I am curious about the trained images you took in CASIA database, was there containing faces with large head-pose and blur?
Or maybe it is just my mistake with some wrong steps.
Hi, you really done a great job! I re-implement your project with C++ code and got almost the same performance as yours in LFW. But when I used it to extract features and test with cosine similarity in CASIA database, the accuracy is only about 80%. The most of the wrong classification caused by large headpose and blurred faces. I am curious about the trained images you took in CASIA database, was there containing faces with large head-pose and blur? Or maybe it is just my mistake with some wrong steps.