Question: Interpretation of resulting confusion matrix

[graulef]

Hey there, I have one more question. After evaluating the network, I came to this resulting confusion matrix:

[eval] accuracy 0.711933
    label 01: 58 [ 58  00  07  24  05  00  00  04  00  00  01  00  00 ]
    label 02: 96 [ 00  96  00  01  01  00  00  02  01  00  00  00  00 ]
    label 03: 62 [ 04  06  62  17  03  00  00  07  01  00  00  00  00 ]
    label 04: 73 [ 13  01  06  73  03  00  00  01  02  00  02  00  00 ]
    label 05: 80 [ 06  02  03  03  80  00  00  00  05  01  01  00  00 ]
    label 06: 43 [ 17  01  09  04  03  43  10  01  01  01  07  00  03 ]
    label 07: 98 [ 00  00  00  00  00  00  98  01  00  00  00  00  00 ]
    label 08: 91 [ 01  01  01  05  01  00  00  91  01  00  00  00  00 ]
    label 09: 93 [ 01  02  00  02  01  00  00  01  93  00  00  00  00 ]
    label 10: 64 [ 04  18  03  01  04  00  00  00  03  64  00  00  02 ]
    label 11: 25 [ 09  00  01  37  02  00  00  25  00  00  25  00  00 ]
    label 12: --
    label 13: 00 [ 100  00  00  00  00  00  00  00  00  00  00  00  00 ]
Training Finished

Now, I know that the results are significantly worse than in the paper, but this is not my question.

My question is, why class 13 (which I believe is the background noise) gets a score of 100 for label 01 in the confusion matrix?

Thanks for your help!

[simonwsw]

To use MaskZero in RNN packages, an additional class need to be reserved for the label of frames that are masked case. Thus this additional class uses 13 (of course only in Torch/Lua) as gesture ID.

Theoretically class 13 shouldn't be shown in the confusion matrix as it's been marked (which is all zeros). However, it shouldn't affect the result.

[graulef]

Okay got it, thank you for the answer!

[mxtx0509]

I miss a question when I want to run the code and I can't solve it ,could you help me? #20