yuezunli/cvprw2019_face_artifacts

Exposing DeepFake Videos By Detecting Face Warping Artifacts

cvprw2019_face_artifacts's Issues

[Errno 20] Not a directory

Why do i always have this error?
Input as below:
python demo.py --input_dir=demo/demo.mp4

RuntimeError: Unsupported image type, must be 8bit gray or RGB image.

Keep running into this issue with every image that I have tried.

Loading checkpoint /home/anubhav/CVPRW2019_Face_Artifacts/ckpt_res50/model
Testing: ./jpg/i056sa-mn.jpg
Traceback (most recent call last):
File "demo.py", line 120, in
run(args.input_dir)
File "demo.py", line 88, in run
prob = im_test(im)
File "demo.py", line 47, in im_test
face_info = lib.align(im[:, :, (2,1,0)], front_face_detector, lmark_predictor)
File "/home/anubhav/CVPRW2019_Face_Artifacts/py_utils/face_utils/lib.py", line 231, in align
faces = face_detector(im, scale)
RuntimeError: Unsupported image type, must be 8bit gray or RGB image.

update the dependencies

hello i have been going through this repository finding that most of the repositories are out-off date and need to update the code for the present libraries

Failed to detect `demo/darpa.mp4`.

I try python demo.py ,but it failed to detect demo/darpa.mp4.

Testing: demo/0000_fake_64.jpeg
Prob: 1.0
Testing: demo/2008_000003.jpg
Prob: -1
Testing: demo/deepFake.mp4
Prob: 1.0
Testing: demo/darpa.mp4
Prob: 1.0

Detection of video

when i detect video，arise finished with exit code -1073741819。how to sovle this problem

Can not load the pretrained model.

Excuse me~ I have downloaded your pre-trained model and want to test the model performance. However, I can not fix this error(as shown below).

My environment is python2.7; TensorFlow 1.3.0;
It seems that the checkpoint does not match the model.meta.

Code for training

Hello,
Can you please share the part of code for training.

RuntimeError: Unable to open /dlib_model/shape_predictor_68_face_landmarks.dat

FP rate is a little high..

The idea in this project is simple but effective. It can really find the deepfake video, but also recongize real video as deepfake video, can you explain why this happen? Do the real videos has warping artifacts?

Tested on res50, don't know how to do the same for other models eg: vgg16

tensorflow.python.framework.errors_impl.InvalidArgumentError: Unsuccessful TensorSliceReader constructor: Failed to get matching files on pretrained_models/vgg_16.ckpt: Not found: pretrained_models; No such file or directory

About detection accuracy

Hello,the generalization ability is quite low, i tested the trained model on UADFV,FF++ and Celeb-DF,but got the accuracy below 50per even lower,so i have the question that does it really work?or something i did wrong?
Hoping for your reply,Thank U.

Generate "fake" face from real one

Hi, I couldn't find in your paper the function that generates the "fake" (not really but you know what I mean) face from the real one. I only managed to find some util funcs.
Can you share this part of the code?
Thank you

Does this model really work? Basically cannot identify non-dataset videos

When I use the model to try to identify the real Obama video, basically all videos are detected with a 95% probability that they are face-changing videos ...

One of the detection videos is as follows：
https://www.youtube.com/watch?v=sHAkDTlv8fA

The clip detection log is as follows：
('path:', '/dlib_model/shape_predictor_68_face_landmarks.dat')
2020-01-19 15:45:07.637407: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.
2020-01-19 15:45:07.637432: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.
2020-01-19 15:45:07.637439: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.
2020-01-19 15:45:07.637445: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX2 instructions, but these are available on your machine and could speed up CPU computations.
2020-01-19 15:45:07.637451: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX512F instructions, but these are available on your machine and could speed up CPU computations.
2020-01-19 15:45:07.637472: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use FMA instructions, but these are available on your machine and could speed up CPU computations.
2020-01-19 15:45:07.815133: I tensorflow/core/common_runtime/gpu/gpu_device.cc:955] Found device 0 with properties:
name: GeForce RTX 2080 Ti
major: 7 minor: 5 memoryClockRate (GHz) 1.545
pciBusID 0000:65:00.0
Total memory: 10.73GiB
Free memory: 10.02GiB
2020-01-19 15:45:07.815162: I tensorflow/core/common_runtime/gpu/gpu_device.cc:976] DMA: 0
2020-01-19 15:45:07.815166: I tensorflow/core/common_runtime/gpu/gpu_device.cc:986] 0: Y
2020-01-19 15:45:07.815189: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1045] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GeForce RTX 2080 Ti, pci bus id: 0000:65:00.0)
Loading checkpoint /home/gky/PycharmProjects/CVPRW2019_Face_Artifacts/ckpt_res50/model
Testing: dataset_test/test/obama_real.mp4
('detecting', 0)
('probs.append:', 0.9868792)
('detecting', 1)
('probs.append:', 0.80856305)
('detecting', 2)
('probs.append:', 0.94188154)
('detecting', 3)
('probs.append:', 0.90303326)
('detecting', 4)
('probs.append:', 0.99767286)
('detecting', 5)
('probs.append:', 0.9614075)
('detecting', 6)
('probs.append:', 0.9595191)
('detecting', 7)
('probs.append:', 0.8166154)
('detecting', 8)
('probs.append:', 0.97614515)
('detecting', 9)
('probs.append:', 0.54254496)
('detecting', 10)
('probs.append:', 0.6386844)
('detecting', 11)
('probs.append:', 0.08990705)
('detecting', 12)
('probs.append:', 0.74948525)
('detecting', 13)
('probs.append:', 0.36588773)
('detecting', 14)
('probs.append:', 0.96228063)
('detecting', 15)
('probs.append:', 0.7279479)
('detecting', 16)
('probs.append:', 0.99728185)
('detecting', 17)
('probs.append:', 0.24376407)
('detecting', 18)
('probs.append:', 0.47520953)
('detecting', 19)
('probs.append:', 0.6742674)
('detecting', 20)
('probs.append:', 0.94513416)
('detecting', 21)
('probs.append:', 0.75601804)
('detecting', 22)
('probs.append:', 0.6526342)
('detecting', 23)
('probs.append:', 0.05435108)
('detecting', 24)
('probs.append:', 0.014932161)
('detecting', 25)
('probs.append:', 0.008214104)
('detecting', 26)
('probs.append:', 0.013158442)
('detecting', 27)
('probs.append:', 0.007968158)
('detecting', 28)
('probs.append:', 0.13961957)
('detecting', 29)
('probs.append:', 0.01184213)
('detecting', 30)
('probs.append:', 0.04152037)
('detecting', 31)
('probs.append:', 0.013649449)
('detecting', 32)
('probs.append:', 0.8725276)
('detecting', 33)
('probs.append:', 0.99986696)
('detecting', 34)
('probs.append:', 0.99989355)
('detecting', 35)
('probs.append:', 0.99965155)
('detecting', 36)
('probs.append:', 0.9992038)
('detecting', 37)
('probs.append:', 0.99040616)
('detecting', 38)
('probs.append:', 0.99737704)
('detecting', 39)
('probs.append:', 0.99870396)
('detecting', 40)
('probs.append:', 0.99959004)
('detecting', 41)
('probs.append:', 0.9452702)
('detecting', 42)
('probs.append:', 0.9321868)
('detecting', 43)
('probs.append:', 0.7133198)
('detecting', 44)
('probs.append:', 0.99842083)
('detecting', 45)
('probs.append:', 0.781696)
('detecting', 46)
('probs.append:', 0.95321435)
('detecting', 47)
('probs.append:', 0.95700437)
('detecting', 48)
('probs.append:', 0.99963474)
('detecting', 49)
('probs.append:', 0.98757046)
('detecting', 50)
('probs.append:', 0.99522686)
('detecting', 51)
('probs.append:', 0.991934)
('detecting', 52)
('probs.append:', 0.99616414)
('detecting', 53)
('probs.append:', 0.99031055)
('detecting', 54)
('probs.append:', 0.98140895)
('detecting', 55)
('probs.append:', 0.9323953)
('detecting', 56)
('probs.append:', 0.9830782)
('detecting', 57)
('probs.append:', 0.9264949)
('detecting', 58)
('probs.append:', 0.971208)
('detecting', 59)
('probs.append:', 0.9561466)
('detecting', 60)
('probs.append:', 0.99475)
('detecting', 61)
('probs.append:', 0.9748002)
('detecting', 62)
('probs.append:', 0.99026966)
('detecting', 63)
('probs.append:', 0.9886033)
('detecting', 64)
('probs.append:', 0.959894)
('detecting', 65)
('probs.append:', 0.9171568)
('detecting', 66)
('probs.append:', 0.9559196)
('detecting', 67)
('probs.append:', 0.9780694)
('detecting', 68)
('probs.append:', 0.9636385)
('detecting', 69)
('probs.append:', 0.965563)
('detecting', 70)
('probs.append:', 0.9263999)
('detecting', 71)
('probs.append:', 0.99452513)
('detecting', 72)
('probs.append:', 0.99809915)
('detecting', 73)
('probs.append:', 0.9980097)
('detecting', 74)
('probs.append:', 0.99830693)
('detecting', 75)
('probs.append:', 0.79650193)
('detecting', 76)
('probs.append:', 0.9734209)
('detecting', 77)
('probs.append:', 0.4492724)
('detecting', 78)
('probs.append:', 0.94296664)
('detecting', 79)
('probs.append:', 0.5423608)
('detecting', 80)
('probs.append:', 0.98778963)
('detecting', 81)
('probs.append:', 0.30337232)
('detecting', 82)
('probs.append:', 0.6011144)
('detecting', 83)
('probs.append:', 0.34220523)
('detecting', 84)
('probs.append:', 0.84168375)
('detecting', 85)
('probs.append:', 0.46755737)
('detecting', 86)
('probs.append:', 0.49161664)
('detecting', 87)
('probs.append:', 0.39203954)
('detecting', 88)
('probs.append:', 0.71996564)
('detecting', 89)
('probs.append:', 0.58421355)
Prob: 0.99316144
Testing: dataset_test/test/obama_real.log
Prob: 0.99316144

training code

Hi,can you share the code for training?

Pretrained Model

Hi~ Thank you for your excellent work on the deepfake detection task. We want to cite your paper and hope to test your pre-trained model on our newly proposed dataset. Can you publish your pre-trained model? thank you very much !

How to warpAffine？

Hi @yuezunli ，Thanks for sharing.I found that it was hard to know how to warpAffine the face to the source image.And I noticed that there was just a random_transform in the proc_img.It seems that it doesn't matter to warpAffine the face (through Gaussian filter) to the source image.Thus I am very intersted that how could you make your own dataset?

Something about AUC

I have tested your paper recently, but its performance is a little different from what you claim.
Could you help me to find the problem？