Extracted faces are too small

[luiscosio]

Expected behavior

Extract faces from frames

Actual behavior

Whenever extracting faces, the process takes too long to start and then the extracted faces are too small.

Example of an extracted face:

This is from an 720 video.

Additionally getting random issues like this

C:\Users\Administrator\faceswap\lib\align_eyes.py:66: RuntimeWarning: overflow encountered in long_scalars
0%| | 9/8456 [22:19<349:09:59, 148.81s/it]eyesCenter = ((leftEyeCenter[0,0] + rightEyeCenter[0,0]) // 2, (leftEyeCenter[0,1] + rightEyeCenter[0,1]) // 2)

Steps to reproduce

Same conda enviroment as local, the only difference is that I'm using AWS with p3 instances

Other relevant information

• Command lined used (if not specified in steps to reproduce): faceswap.py ...
• Operating system and version: Windows 2016 on AWS
• Python version: 2.7, 3.5, 3.6.4, ...
• Faceswap version: latest on master
• Faceswap method: GPU
• Other related issues: Face state detection & categorization? #123, Small error in lib/cli.py get_faces(self, image)  #124...

absl-py 0.2.0
astor 0.6.2
bleach 1.5.0
certifi 2018.4.16 py36_0 conda-forge
click 6.7
cycler 0.10.0
Cython 0.28.2
decorator 4.3.0
dlib 19.9 np111py36_0 conda-forge
face-recognition 1.2.2
face-recognition-models 0.3.0
ffmpeg 4.0 hf48ec3a_0 conda-forge
ffmpy 0.2.2
ffprobe 0.5
gast 0.2.0
grpcio 1.11.0
h5py 2.7.1 py36he54a1c3_0
hdf5 1.10.1 vc14_2 [vc14] conda-forge
html5lib 0.9999999
icc_rt 2017.0.4 h97af966_0
intel-openmp 2018.0.0 8
Keras 2.1.6
keras-vggface 0.5
kiwisolver 1.0.1
libpng 1.6.34 vc14_0 [vc14] conda-forge
Markdown 2.6.11
matplotlib 2.2.2
mkl 2018.0.2 1
mkl_fft 1.0.1 py36h452e1ab_0
mkl_random 1.0.1 py36h9258bd6_0
networkx 2.1
numpy 1.14.3
numpy 1.11.3 py36h4a99626_4
opencv-python 3.4.0.12
pathlib 1.0.1 py36_0 conda-forge
Pillow 5.1.0
pip 10.0.1 py36_0
protobuf 3.5.2.post1
pyparsing 2.2.0
python 3.6.5 h0c2934d_0
python-dateutil 2.7.2
pytz 2018.4
PyWavelets 0.5.2
PyYAML 3.12
scandir 1.5 py36_0
scikit-image 0.13.1
scipy 1.1.0
setuptools 39.1.0 py36_0 conda-forge
six 1.11.0 py36h4db2310_1
tensorboard 1.8.0
tensorflow-gpu 1.8.0
termcolor 1.1.0
tqdm 4.23.3
vc 14 h0510ff6_3
vs2015_runtime 14.0.25123 3
Werkzeug 0.14.1
wheel 0.31.0 py36_0 conda-forge
wincertstore 0.2 py36h7fe50ca_0
zlib 1.2.11 vc14_0 [vc14] conda-forge

[luiscosio]

Updated the environment and changed conda to virtualenv:

absl-py                 0.2.2
astor                   0.6.2
bleach                  1.5.0
click                   6.7
cloudpickle             0.5.3
cycler                  0.10.0
dask                    0.17.5
decorator               4.3.0
dlib                    19.10.9
face-recognition        1.2.2
face-recognition-models 0.3.0
ffmpy                   0.2.2
gast                    0.2.0
grpcio                  1.12.1
h5py                    2.8.0
html5lib                0.99999
Keras                   2.1.6
kiwisolver              1.0.1
Markdown                2.6.11
matplotlib              2.2.2
networkx                2.1
numpy                   1.14.3
opencv-python           3.4.1.1
pathlib                 1.0.1
Pillow                  5.1.0
pip                     10.0.1
protobuf                3.5.2.p
pyparsing               2.2.0
python-dateutil         2.7.3
pytz                    2018.4
PyWavelets              0.5.2
PyYAML                  3.12
scandir                 1.7
scikit-image            0.14.0
scipy                   1.1.0
setuptools              39.2.0
six                     1.11.0
tensorboard             1.8.0
tensorflow-gpu          1.8.0
termcolor               1.1.0
toolz                   0.9.0
tqdm                    4.23.4
Werkzeug                0.14.1
wheel                   0.31.1

Still having the same issue, even though face_recogntion works just fine.

[torzdf]

This looks to be almost definitely a dlib issue. What happens if you debug landmarks?

Which Cuda and CuDNN versions are you using?

[luiscosio]

@torzdf The landmarks are a mess.

CUDA 8 & 9, cuDNN 6 & 75 and NVidia Driver 385.54.

But I just found out it was related to using TESLA V100

Exact same stack in Pascal works like a charm.