Low results

[FraLotito]

Hi!

Why are the results so low with respect to the paper? Have you found an explanation?

Thanks

[zhuyijie88]

+1, the reported results are worse than these in the paper by a large margin.

[aphalak]

I think the model implementation is incorrect for the voting module layers.
In the paper, the weights are shared such that each seed feature is passed forward separately, so there should be a for loop of forward passes per seed feature, whereas in the implementation here, there is a single forward pass over the FullyConnected layer.

[qq456cvb]

@aphalak Though each seed feature is passed forward separately in the paper, the weights are shared so we could batch them together and do a single forward. That is not the reason of low results. I am also figuring out what is happening and it is some what related to Issue #9 . Since all computations are gathered in the graph, when there is no "positive" vote, the code continues in vain, making training inefficient.

[619862306]

Could you please give me a copy of the script file of tf_nms3d_compile.sh?