![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1440116%2F0b82c4e15c56ac09e24b3c58831dda75%2F10k_avg.png?generation=1583921611945437&alt=media\")
\nversus facehq (1024 x 1024 resolution)\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1440116%2F4d429be27df8d5e4a1bcab0d8a08d0c1%2F1khq_avg.png?generation=1583921665784387&alt=media\")
Paperswithcode is a great website to get the code associated with articles.
\n\nSearch: https://paperswithcode.com/search?q_meta=&q=deepfake
\n\nBest (matching) results (no order):
\n\nPaperswithcode is a great website to get the code associated with articles.
\n\nSearch: https://paperswithcode.com/search?q_meta=&q=deepfake
\n\nBest (matching) results (no order):
\n\nOne interesting idea that caught my attention: computing azimuthally averaged radial profile from the Unmasking DeepFake with simple Features paper.
\n\nHere is the function (source: https://github.com/cc-hpc-itwm/DeepFakeDetection/blob/master/Experiments_DeepFakeDetection/radialProfile.py):
\n\n```\nimport numpy as np
\n\ndef azimuthalAverage(image, center=None):\n \"\"\"\n Calculate the azimuthally averaged radial profile.\n image - The 2D image\n center - The [x,y] pixel coordinates used as the center. The default is \n None, which then uses the center of the image (including \n fracitonal pixels).
\n\n\"\"\"\n# Calculate the indices from the image\ny, x = np.indices(image.shape)\n\nif not center:\n center = np.array([(x.max()-x.min())/2.0, (y.max()-y.min())/2.0])\n\nr = np.hypot(x - center[0], y - center[1])\n\n# Get sorted radii\nind = np.argsort(r.flat)\nr_sorted = r.flat[ind]\ni_sorted = image.flat[ind]\n\n# Get the integer part of the radii (bin size = 1)\nr_int = r_sorted.astype(int)\n\n# Find all pixels that fall within each radial bin.\ndeltar = r_int[1:] - r_int[:-1] # Assumes all radii represented\nrind = np.where(deltar)[0] # location of changed radius\nnr = rind[1:] - rind[:-1] # number of radius bin\n\n# Cumulative sum to figure out sums for each radius bin\ncsim = np.cumsum(i_sorted, dtype=float)\ntbin = csim[rind[1:]] - csim[rind[:-1]]\n\nradial_prof = tbin / nr\n\nreturn radial_prof\n```
", "rawMarkdown": "One interesting idea that caught my attention: computing azimuthally averaged radial profile from the **Unmasking DeepFake with simple Features** paper. \n\nHere is the function (source: https://github.com/cc-hpc-itwm/DeepFakeDetection/blob/master/Experiments_DeepFakeDetection/radialProfile.py): \n\n```\nimport numpy as np\n\ndef azimuthalAverage(image, center=None):\n \"\"\"\n Calculate the azimuthally averaged radial profile.\n image - The 2D image\n center - The [x,y] pixel coordinates used as the center. The default is \n None, which then uses the center of the image (including \n fracitonal pixels).\n \n \"\"\"\n # Calculate the indices from the image\n y, x = np.indices(image.shape)\n\n if not center:\n center = np.array([(x.max()-x.min())/2.0, (y.max()-y.min())/2.0])\n\n r = np.hypot(x - center[0], y - center[1])\n\n # Get sorted radii\n ind = np.argsort(r.flat)\n r_sorted = r.flat[ind]\n i_sorted = image.flat[ind]\n\n # Get the integer part of the radii (bin size = 1)\n r_int = r_sorted.astype(int)\n\n # Find all pixels that fall within each radial bin.\n deltar = r_int[1:] - r_int[:-1] # Assumes all radii represented\n rind = np.where(deltar)[0] # location of changed radius\n nr = rind[1:] - rind[:-1] # number of radius bin\n \n # Cumulative sum to figure out sums for each radius bin\n csim = np.cumsum(i_sorted, dtype=float)\n tbin = csim[rind[1:]] - csim[rind[:-1]]\n\n radial_prof = tbin / nr\n\n return radial_prof\n\n```", "votes": null }, { "id": "768859", "postDate": "03/11/2020 09:13:52", "content": "I think not very separable with low resolution picture \nBelow is average 10000 image from https://www.kaggle.com/dagnelies/deepfake-faces\n\nversus facehq (1024 x 1024 resolution)\n
One interesting idea that caught my attention: computing azimuthally averaged radial profile from the Unmasking DeepFake with simple Features paper.
\n\nHere is the function (source: https://github.com/cc-hpc-itwm/DeepFakeDetection/blob/master/Experiments_DeepFakeDetection/radialProfile.py):
\n\n```\nimport numpy as np
\n\ndef azimuthalAverage(image, center=None):\n \"\"\"\n Calculate the azimuthally averaged radial profile.\n image - The 2D image\n center - The [x,y] pixel coordinates used as the center. The default is \n None, which then uses the center of the image (including \n fracitonal pixels).
\n\n\"\"\"\n# Calculate the indices from the image\ny, x = np.indices(image.shape)\n\nif not center:\n center = np.array([(x.max()-x.min())/2.0, (y.max()-y.min())/2.0])\n\nr = np.hypot(x - center[0], y - center[1])\n\n# Get sorted radii\nind = np.argsort(r.flat)\nr_sorted = r.flat[ind]\ni_sorted = image.flat[ind]\n\n# Get the integer part of the radii (bin size = 1)\nr_int = r_sorted.astype(int)\n\n# Find all pixels that fall within each radial bin.\ndeltar = r_int[1:] - r_int[:-1] # Assumes all radii represented\nrind = np.where(deltar)[0] # location of changed radius\nnr = rind[1:] - rind[:-1] # number of radius bin\n\n# Cumulative sum to figure out sums for each radius bin\ncsim = np.cumsum(i_sorted, dtype=float)\ntbin = csim[rind[1:]] - csim[rind[:-1]]\n\nradial_prof = tbin / nr\n\nreturn radial_prof\n```
", "votes": null, "replies": [ { "id": 768859, "author_name": "nyleve", "author_url": "", "post_date": "03/11/2020 09:13:52", "content": "I think not very separable with low resolution picture \nBelow is average 10000 image from https://www.kaggle.com/dagnelies/deepfake-faces\n\nversus facehq (1024 x 1024 resolution)\n