Some information I want to share regarding accuracy

[Dong-JinKim]

I am 20153080 I want to share some mistakes i have done.

1) about LayerId we need to extract CNN activation from the correct layer

if the LayerId=19 we need to extract from 'ReLU7' layer.not 'FC7' layer.

Even though activation from 'FC7' worked quite well and gives us reasonable accuracy, activation from 'ReLU7' provided even more accurate predictions

2) normalization during Fisher encoding As TA told us, normalization is very important during Fisher encoding

I found that the result with power and L2 normalization outperformed the result without normalization in about 1~2% accuracy

I hope everyone of us finish this PA safely. Thank you

[BAILOOL]

So you mean we need to extract descriptors from layer = LayerId+1?

[cvpa3]

Actually, the output activation of LAYER_ID of a network exactly corresponds to LAYER_ID+1 of the result from vl_simplenn(). So you must extract ReLU7 activations in this PA.