Kaggle IDR&D Antispoofing Challenge

The competition organized on the platform DataSouls has the objective to develop the algorithm detection of spoofing attacks in face recognition.

Hardware

• GPU: 1xTesla K80

Data format

The images are represented in the form of frames in video sequence having the following attributes: label, path, video.

Solution

The solution represents:

• model: ResNet50
• validation metrics: minC = FP / (FP + TN) + 19*FN / (FN + TP)
• augmentations: RandomHorizontalFlip, RandomResizedCrop, ColorJitter
• optimizer: Adam
• loss: BCELoss
• batch size: 64