There are some problems in the implementation of 'recurrent_hourglass.py'.

[sunyasheng]

In the implementation of 'recurrent_hourglass.py'. There is only oneConvGRUCell

self.convgru_cell = ConvGRUCell

But it seems that there are supposed to be n_levels ConvGRUCell, each of which tackles one type of resolution feature maps. As for the recurrent part, I think you can just maintain the hidden state of GRUCell and let the input at next timestep come in.

That's my understanding of the paper Instance Segmentation and Tracking with Cosine Embeddings and Recurrent Hourglass Networks. If you think in another way, let me know.