How Many samples Needed to training the Conv3D?


Hi Everyone,

I am currently working in human activities classification based on video sequence (16 Frames model input) and Dataset size is 300. If I train the simple five conv3d layer (Relu->Max pooling, output->softmax) with this sample, the model is overfitting. I have tried temporal data-augmentation and increased the sample into 1000, but the model is overfitting.

Please give me a suggestion for data-augmentation and training the conv3d model.