Reviews: Deep Learning for Precipitation Nowcasting: A Benchmark and A New Model

This paper introduces TrajGRU, an extension of the convolutional LSTM/GRU. Contrary to convLSTM/GRU, TrajGRU aims at learning location dependant filter support for each hidden state location. TrajGRU generates a flow field from the current input and previous hidden state and then warp the previous hidden states through bilinear sampling following this flow field. Author evaluate their proposal on a video generation on two datasets, MovingMNIST having 3 digits at the same time and HKO-7 nowcasting dataset, where TrajRU outperforms their convolutional counterpart. Few specific question/remarks: Did you compare TrajGRU with ConvGRU having a larger support than just a 5x5 kernels?