fidler-lab / curve-gcn Goto Github PK

Hi I AM TRYING TO GET THE CODE. BUT I AM UNABLE TO DO SO

It seems there is no interactive mode code.

We have predictions from model as new position of the vertices (x_new, y_new) for each node. In the accuracy loss do we create a segmentation mask (of 1s and 0s) from these predicted points and take L1 norm with ground truth mask? If so can you please explain in short how it is done and why this is not differentiable? Thanks.

Hi,

you have written in your paper that the code is publicly available, but when I contacted you, I got the response that the code is not public. Could you please explain this?

In the paper, you have wirtten the algorihtm about how to train the interactiveGCN, but I'am confused to the "Do not stop gradients".
In algorithm 1, it look's like InteractiveGCN is the inference procedure and the TrainInteractiveGCN is the train procedure. Can you explain the "Do not stop gradients" more clearly?

Hi,

First of all, thanks for open sourcing such a great piece of work. Currently, I cannot access to the "signup" page to download the code. Could you please check this issue?

Any response will be helpful. Thank you again for your wonderful work.

Hi, @arieling, @amlankar. Thanks for your work. I have some questions about the InteractiveGCN part in your paper. You said in your paper that "we now append two additional dimensions to the corrected node's input feature and keep the network's architecture the same as the original GCN"
and the representing as the followed furmula:

I have 3 questions about the interactive mode that you said in your paper:
1. I checked the code that you public code, at the forward GCN infer part:
gcn_pred = self.gnn[i].forward(input_feature, adjacent)
as the code shown, the dimention of input_feature is (batch_size, 40, 258), the dimention of adjacent is (batch_size, 40, 40). Does the representing furmula means when do interactiveGCN inferred process, the input_feature dimention should be (batch_size, 40, 260)? And the extra 2 dimentions are the (∆x, ∆Y). Does it right or not? If i'm wrong, can you tell me how the extra dimention of (∆x, ∆Y) should be presented when do interactive inferring part?
2. You said the GCN in your InteractiveGCN "asme as original GCN", the GCN forward in your public code is combined by two GraphConvolution and six GraphResConvolution operations
, that means in your InteractiveGCN mode, the architecture is the same as the auto mode or just using a few of GraphConvolution operation to implement your interactiveGCN except the input_feature added two extra dimentions, in other words, does your interactiveGCN input_feature is (batch_size, 40, 260) dimentions?
3. The last problem also confused me for a long time. In your InteractiveGCN mode, Does the input image is fed in to the autoGCN inferring part firstly, and then feed the output of the autoGCN in the InteractiveGCN to get the final InteractiveGCN results? Because you said "To make our model interactive, we tarin another GCN " in your paper. Can you tell me what's the difference between interactiveGCN and the original GCN more? I think if you only train the interactiveGCN without the initilization of autoGCN's parameter, the interactiveGCN's results may be not good than the autoGCN's results.
Appretiative for your reply ~^_^

Have you evaluated the performance on high quality masks like Davis?

https://davischallenge.org/challenge2019/index.html

Hello,

We have been waiting public code about your excellent job's result.

When can we get ?

I would like to deploy this interactive annotations on images with overlapping chromosomes.
Does the model supports annotation of overlapping masks?

hi, @arieling , Thanks for your great work. And when will you plan to publish the online tools and interactive mode of cvpr19 paper "curve-gcn"? I'm very interested in your work^_^. Appreciative for your reply.

They are all with CityScapes.

1. Like this figure, let's see the light blue curve named 'Our T=1'. In my opinion, you will correct all the objects until they are all with IoU 100, then why the AVG IoU is just below 90 (it should be 100)?

2. Another question, why the start IoU is about 77? In the table 1 you have IoU 73+ w/o correction.
   

3. When you calculate the AVG correction, do you calculate it in the way in PolygonRNN? It calculate every category's mean correction number(total_category_correction/total_category_instance_num, no matter the instance is corrected or not), and sum them then divide by 8 (8 category). Is my understanding correct？

Dear Team. I have requested for the access to the code for my research using my university email id 2 weeks ago and yet I haven't received the link for the source code. It would be great if anyone could look into this. :)

It seems like the prediction scipt is incorrect.
Moreover, could you please tell me more details on how to use manual mode? I modify the gnn-active-spline-diff-render.json and only get some png result.

I aslo get some warrning:

/home/yan/workspace/curve_gcn_release/code/active_spline/ActiveSplineDiffIOU.py:5: UserWarning:
This call to matplotlib.use() has no effect because the backend has already
been chosen; matplotlib.use() must be called before pylab, matplotlib.pyplot,
or matplotlib.backends is imported for the first time.

Hi there~
You said below in ablation study
Boundary Pred adds the boundary prediction branch to our CNN.
Does the 'Boundary Pred' mean below
In order to help the model see image boundaries, we supervise two additional branches, i.e. an edge branch and a vertex branch, on top of the CNN encoder’s feature map F c , both of which consist of one 3 × 3 convolutional layer and one fully-connected layer. These branches are trained to predict the probability of existence of an object edge/vertex on a 28 × 28 grid.
I am a little confused, because the GCN also predict the boundary.

As title

Hi,

It seems that the dataset can only be downloaded by browser. But it's too large, the server finally disconnected after a long time, and it cannot support resuming from break point.

Any advice? Or will you support a direct download link?

Hi, I was interested in the claims in the paper that Curve-GCN also works reasonably well on out-of-domain datasets. How do I test on my own dataset? Does it work by just changing the data directory and the "class_filter" arguments in the experiment json file? I'm assuming this step would be easier with the interactive mode?

Thank you.

First of all, thanks for your excellent work.
However, when I wanted to sign up in your website(Curve-GCN Code Signup (toronto.edu)), I met a problem. I can open your website but I failed when I press the submit button. What should I do to solve the submission problem? Or is there another way to get your full-version curve-gcn code?
Looking forward to your reply.

Thank you for your great great work about polygon infer by GCN! I have learned a lot in you paper.

I have signed for this code but got reply：
We are currently not sharing the code, we will send a notification as soon as it is available again. In the meantime, we are sharing the Polygon-RNN++ code (www.cs.toronto.edu/polyrnn) for research purposes. Thank you for your patience!

So when will the code be shared? Could I request the old version code for helping me with my graduate project(academic)??? Or only the network architecture (like this part: https://github.com/fidler-lab/curve-gcn/code/Models).

I'll appreciate it if you could shared a old (part) version of GCN regression, many thanks. :)

Does anybody get the code?