Unet3D Based Next Frame Prediction

The concept of next-frame prediction, which is predicting the subsequent frames using historical frames’ spatial and temporal properties, is indispensable in computer vision. There are various application of frame prediction such as predicting a future event in autonomous vehicles, predicting patient falls in biomedical engineering, and reducing the amount of data transmitted in video transmission. Deep learning applications in this field are the focus of the most effective methods. Especially CNN-LSTM, Convolutional LSTMs, and GAN-supported deep learning methods are very common. This study proposes the inflated 3D Unet encoder-decoder model, which is not yet used for the next-frame prediction problem. The proposed model predicts both the next frame and the subsequent frames. Experimental results have shown that the proposed method gives better results than CNN-LSTM and Convolutional LSTMs.