Code of GCN-S part about learn-to-cluster HOT 21 OPEN

GCN-S is available now. Feel free to have a try. @LZDSJTU @CaiQiuYu @AruniRC @zsz00

from learn-to-cluster.

@LZDSJTU We will try to release as much code as possible before the conference.

from learn-to-cluster.

@AruniRC Thanks for checking our progress. Since the difference between GCN-S and GCN-D is not very significant, the priority of cleaning GCN-S is lower than our other projects currently. Actually, the model of GCN-S is available at dsgcn model. By setting the reduce_method to no_pool and stage to seg, it becomes the GCN-S model. One missing thing is introducing random seed in data processor to decide the segmentation ground-truth, which can be implemented through writing a cluster_seg_processor as cluster_det_processor. It is always in our schedule and we will find some time to work on that.

from learn-to-cluster.

@linjiawen2019 Thanks for reporting! We will fix it through PR #44.

from learn-to-cluster.

@linjiawen2019 The PR is merged. Feel free to have a try.

from learn-to-cluster.

@jxyecn We use the frames extracted by YTF dataset. Checkout the second or third download link in YTB official page. Particularly, as stated in experimental settings, we use 14,653 frames for training and 140,629 frames for testing. The related data and pretrained model have been uploaded to GoogleDrive and OneDrive to reproduce the result. You can also find it in README of the dev branch.

from learn-to-cluster.

Thank you for your contribution.

Additionally, could you provide the list for splitting MS-Celeb-1M dataset?

from learn-to-cluster.

@LZDSJTU The list has been uploaded to GoogleDrive and OneDrive. You can also find it in README of the dev branch.

from learn-to-cluster.

Hello, I saw you evaluated your method with YTF dataset, I wonder how to generate this dataset?
How to sample images from YTF's original video frames, can you provide more detailed information?

from learn-to-cluster.

@yl-1993 Thanks for your reply

from learn-to-cluster.

@LZDSJTU We will try to release as much code as possible before the conference.
期待尽早开源GCN-S部分的代码，好像从论文上看，GCN-D+GCN-S性能提升比起只用GCN-D并不是很多么？@yl-1993

from learn-to-cluster.

@CaiQiuYu GCN-S is devised to improve the precision of the proposals. Since the precision of clustering results after GCN-D is over 95, there is no much space for improvement. In ablation study, we show that GCN-S can be used as an independent module to combine with other methods.

from learn-to-cluster.

Is there any news about publishing the training code?

from learn-to-cluster.

@LZDSJTU Thanks a lot for checking our progress. The training code will be available after PR #6 is merged.

from learn-to-cluster.

Hi @yl-1993 , is there any plan of releasing the code of the GCN-S ?

from learn-to-cluster.

Hi @yl-1993 , is there any plan of releasing the code of the GCN-S ?

from learn-to-cluster.

@zsz00 Thanks for checking. I'm busy preparing a submission and currently may not have much time for this project. For those who want to have a try, I provide a code snippet for GCN-S. The snippet can not run directly, but includes the data processing and loss function used in GCN-S.

from learn-to-cluster.

Thanks for your reply

from learn-to-cluster.

Hi all, GCN-S will be available after PR #31 is merged.

from learn-to-cluster.

@yl-1993 Hello, I would like to ask some questions about testing GCN-S with my own dataset which has no id labels. There's something wrong when I run "test_cluster_seg_ms1m.py", it occurred an error that "TypeError: 'NoneType' object is not subscriptable " in cluster_seg_processor.py. Here the 'NoneType' object refers to dataset.idx2lb
When building processor for GCN-S, a function "build_graph" in class ClusterSegProcessor will compute labels for data, my question is how to compute labels when testing data have no labels (which means dataset.idx2lb is None)?
I'll appreciate it if you could answer my question!! Thank you!

from learn-to-cluster.

@linjiawen2019 Thanks for reporting! We will fix it through PR #44.

Thank u so much!! Really helpful to me

from learn-to-cluster.