How can i make annotations of objects interactively using humans-in-the-loop? about polyrnn-pp

[amlankar]

Hi @Jacoobr,

1. Thanks for trying it out! You are correct, when you correct the vertex at time step T, the vertices before that do not get corrected. In our tool, when you drag a vertex, the vertices that get coloured red will not move. This is why the right method of interactive annotation is to move the first vertex that is wrong in the direction of RNN prediction to the right point (the direction is specified by the green line leaving from the green point in our tool).

2. We are working on preparing a training code release that is clean and modular. We will update on the repository as soon as it is done!

[amlankar]

Hi, if you're asking for the code for this, then it is in the works right now.

If you are asking theoretically, then the idea is to remember which time step the correct was made in, and force-feed the RNN the vertex at that time step, instead of letting it take its own output. Does that make sense?

[Jacoobr]

@amlankar Thanks for your reply, I can see the result when I correct one vertex that was produced not good by the RNN on your great online tools, then the whole vertexts except the vertext that I corrected will be reproduced (predicted) by the model nicely. But there are two questions as follows confused me: 1 ) I think when the vertex at time step t which corrected by me, then the next whole vertexs after time step t will be predicted once again depend on the vertext that I corrected. But the vertext before time step t should not be re-predicted because the lstm RNN net work is a sequence structure. In other words, the vertext corrected at time step t can't donate to the vertexts before that time step. How do you think of my thought? 2) Would you mind provide me with the training network? I plan to train my own model on MRI medical images data set about MRI segmentation task research. To be honest, I'm a newbie and I have taken a few days on trying to write the training network according to your CVPR 2018 paper. But I get nowhere about it. Your help will help me a lot with my work . This is my email address [email protected]. Look forward to your favourable reply. Thanks.

[yshvrdhn]

Hi any timeline as to when will the correction code, as well as training code, will be made available? Thank you for releasing the inference code.

[amlankar]

It'll be released by the end of the summer

[lucasjinreal]

@amlankar Wish for your work and the release of training codes. This work really marvelous!

[amlankar]

We have released the code now! Checkout https://github.com/fidler-lab/polyrnn-pp-pytorch