anson0910/CNN_face_detection

Hi,about the train_cal accuracy and loss!

Closed this issue · 4 comments

I train the 12_cal by myself but i find the loss is about 2.3 to 2.7 and the accuracy is about 0.3 after iteration 10W, is it right?

transform_param {
#crop_size: 12
mean_value: 104
mean_value: 117
mean_value: 123
#mirror: true # maps [0 255] to [0 1]
}

is the mean_value relative to the following file?

I0411 17:56:35.388991 17623 compute_image_mean.cpp:112] Number of channels: 3
I0411 17:56:35.388999 17623 compute_image_mean.cpp:117] mean_value channel [0]:114.887
I0411 17:56:35.389019 17623 compute_image_mean.cpp:117] mean_value channel [1]:129.801
I0411 17:56:35.389024 17623 compute_image_mean.cpp:117] mean_value channel [2]:154.07

Beside, i find that if do not use the crop_size: 12 and mirror: true the accuracy is about 0.8 for the 12_cal.

is the mean_value relative to the following file?
Yes!

Beside, i find that if do not use the crop_size: 12 and mirror: true the accuracy is about 0.8 for the 12_cal.
Great! The mirror attribute should not be used for calibration nets, that should be a mistake of mine.

about the 12_cal In the paper it says the sliding window is:

total:5341.8
after 12-net:426.9
after 12-calibration-net:388.7

but in your code there is no window reduce after 12-calibration-net, is there anything different?

I think in the paper, it meant after 12-calibration-net, then after local-NMS, the number of windows should reduce.
The number of windows should not change after calibration.