Inferencing without face detector leads to very low performance

mapooon / SelfBlendedImages

[CVPR 2022 Oral] Detecting Deepfakes with Self-Blended Images https://arxiv.org/abs/2204.08376

Other

211 stars 34 forks source link

Inferencing without face detector leads to very low performance #14

Open deepfakeLearnerHi opened 2 years ago

deepfakeLearnerHi commented 2 years ago

Hi, I have my face dataset ready and wish not to perform duplicate face detection during inference as it brings huge time cost. However, when I commented out the face detection code and directly ran the inference on my cropped faces (using FFmpeg and dlib) from FF++, it seems that the performance dropped a lot. I wonder if there is a solution to it.

githuboflk commented 1 year ago

@tianyw23
Can you provide the relevant code? I was in a similar situation.

Pudge-tao commented 1 year ago

me too，I changed the face detector by Centerface and test on FF++ that got a terrible auc

mapooon commented 1 year ago

Hi, The performance of the classifier strongly relies on the preprocess, e.g., cropping. We recommend you use the same preprocess as we provide in this repository.

leobryan commented 1 year ago

me too，AUC shows around 0.50 for all cross datasets, wonder why