Closed aravadikesh closed 1 day ago
Weights provided by repo owners are unavailble and had to be sourced externally from reviewers. Unable to get comprehensive output from video dataset, will attempt pre-processing of data to see if better results can be produced.
Found model weights at the following link https://data.lip6.fr/cadene/pretrainedmodels/. Attempted to run XceptionNet on a small dataset sourced from SDFVD https://data.mendeley.com/datasets/bcmkfgct2s/1. Model runs successfully but the output is incomprehensible. It detects all input videos as real videos, even very obvious deepfakes. Modified the pipeline to check for obvious root causes, but was unable to find any.
Also tried to reach out to the repo owners but received no response yet. Will try for a little longer to locate the issue, else will move forward with the alternative model found by Prit.
Evaluating viability of following Xception model on local:
https://github.com/ondyari/FaceForensics/tree/master/classification