You are on page 1of 2

I’m using 5 Smartphones (C1, C 2, C3, C4 and C5).

➢ Step 1: As we agreed, I have created 5 CNN networks (one for each device, one vs all),
using a cropped block (224x224). The output of each CNN is a feature map. (1000
features × 140729 observations).
➢ Step 2: I applied PCA for each feature map and tested the performance with different
PCA components.
➢ Step 3: Create 5 GMMs (one for each device, positive GMM only).
➢ Step 4: I know I should use the testing data, but I just went to see the performance
training data, so I have compared it with the 5 positive GMMs using the probability
density function for Gaussian mixture distribution (pdf). The value corresponds to the
smartphone which has been used to record the video. However, the accuracy was very
low (25% to 30%); the reason could be that hundreds of rows in the score vector are
shown as zero (row values equal zero, so there is no max value see the below example,
each column in the score represents the score for one GMM (class), see the PDF file).
Let me have your feedback.
Thanks.
GMM1 GMM2 GMM3 GMM4 GMM5
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0

You might also like