Skip to content

Project 3 of Trends and Applications of Computer Vision course of UniTN

Notifications You must be signed in to change notification settings

RickyZi/taaocv-project

 
 

Repository files navigation

Project 3 of Trends and Applications of Computer Vision course of UniTN

In this project we wants to asses various technique to detect a deepfake and to try them on few different deefake generators (FSGAN, DeepFaceLive, GHOST).

Students

  • Giovanni Lorenzini
  • Diego Planchenstainer
  • Riccardo Sassi
  • Riccardo Ziglio

Challenges

The human suspected to be a deepfake is required to perform some specific actions, such as:

  • Head rotation: move the head with strong rotation as 90° horizontally or face up towards the ceiling. This aims at assess whether the model can generate unseen views.
  • Tongue out: stick a portion of the tongue out. If the segmentation network is capable of detecting the tongue, it will be ignored and thus displayed, otherwise the face swapper will delete it.
  • Hand on face: hover hand and finger in front of the face, either occluding it partially or completly. The objective is to evaluate the performance of the segmentation network.
  • Close up: move the face close to the camera. This step wants to assess whether the face detector is capable of detecting only part of face and the resolution of the face.
  • Standup: stand up in order to hide the face from the camera. This step wants to assess whether the face detector is capable of re-detecting a face that was hidden/gone out of the scene.
  • Poke cheek: poke cheek to induce morphological changes in the face. These changes should be sensed by the landmark detector and face allignment networks.
  • Expression: perform different expression. These changes should be sensed by the landmark detector and face allignment networks.
  • Speaking: speak to generate lips movement. These changes should be sensed by the landmark detector and face allignment networks.

Dataset

We used some videos of ourself performing the above actions and videos downloaded from the internet with the celebrity that we want to impersonate.

Deepfake Generators

We used the following deepfake generators:

  • FSGAN:
    • uses a small video for source face;
  • DeepFaceLive:
    • use pretrained models from DeepFaceLab;
    • in real time;
  • GHOST:
    • uses a single image for source face.

Results

FSGAN GHOST DeepFaceLive
Head rotation 0.5 2.8 4.8
Tongue out 4.2 2 0.5
Hand on face 1.8 2.5 5.3
Close up 5 4 7.8
Standup 0 7.8 5
Poke cheek 0 5.3 5
Expression 7 7.3 8.5
Speaking 6.5 6 9
Average 3.1 4.7 5.7
Realism 4.8 7.5 9
Lighting conditions 2 6 9

About

Project 3 of Trends and Applications of Computer Vision course of UniTN

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Jupyter Notebook 100.0%