I just use the first split of the UCF101,and follow the preprocessing and data augmentation of I3D paper,
and normalization is easy ,just follow [(x - 127) / 128 + 1] / 2 , where x is pixel value 。But test accuracy is far below than train accuracy , train accuracy can attach above 90%
I just use the first split of the UCF101,and follow the preprocessing and data augmentation of I3D paper, and normalization is easy ,just follow [(x - 127) / 128 + 1] / 2 , where x is pixel value 。But test accuracy is far below than train accuracy , train accuracy can attach above 90%