Problems on Mis-Recognitions on Non-Living Objects

[TMaysGGS]

Hi @ipazc ,

Thanks very much for sharing your work and this version of mtcnn is really precise compared with some other versions.

But I found that there are some cases that the detector will always detect some non-living objects as human face, such as mugs with patterns, down jackets, parts of coats with metal buttons and hands holding a cellphone. I tried adding some preprocessing steps before the image goes into the network but did not get the results imporved significantly. And I extracted the confidence for those pics with non-living objects, they can get a confidence as high as 0.9+ after going into the ONet, which makes adjusting the threshold for ONet not work. So I wonder if you or anyone have met similar situations, and what modifications can be applied here to eliminate the error.

Below are some of the mis-recoginized samples.

Any advice would help. Thanks.

[ipazc]

Hi, you could try adjusting the steps_threshold parameter in the MTCNN constructor (which by default is [0.6, 0.7, 0.7]) to better fit your images distribution.

from mtcnn import MTCNN

detector = MTCNN(steps_threshold=[0.6, 0.7, 0.7])

Those values are the default ones given by the MTCNN author, but might not be the best values for all the scenarios.

[TMaysGGS]

Hi, you could try adjusting the steps_threshold parameter in the MTCNN constructor (which by default is [0.6, 0.7, 0.7]) to better fit your images distribution.

from mtcnn import MTCNN

detector = MTCNN(steps_threshold=[0.6, 0.7, 0.7])

Those values are the default ones given by the MTCNN author, but might not be the best values for all the scenarios.

Hi, @ipazc ,

Thank you for your reply.

I tried changing the values of the thresholds, and many of those false positives have a confidence as high as 0.8, 0.9 for RNet & ONet, which makes it not work.

Also, I tried freezing your network parameters and add "Dense - PReLU - Dense - Classifier" at the end of the network main branch, which makes ONet have 4 parrallel outputs, one for filtering those FPs. And it is still not working (maybe because I do not have enough samples).

PS: I did not retrain your network as a whole. Because I wrote the training code myself but for now I still cannot get a relatively good PNet. So I am afraid of breaking the good weights that have been balanced already in your network. Here is my training code: https://github.com/TMaysGGS/MTCNN-Keras.