kuangjuihsu / deepco3 Goto Github PK

Hi,

Great work and very impressive results.
I have three problems related to the paper and hope you could help me solve this:

1. Based on your experience, how many images of a certain class are required to train DeepCO3 to converge well?
   In your provided dataset, the number of images of cows, sheep and horses are more than that of trains, air-plane, bus. Correspondingly, results on categories with more images seem better than others.
   (I suppose we train different DeepCO3 for different classes instead of mixing them.)

2. If i understand correctly, the whole pipeline of DeepCO3 is fully un-supervised, isn't it?

3. You mentioned in Sec. 3.3 you used MCG for object proposals which is unsupervised. However, MCG seems to use BSD dataset to get a good setup for parameters. Do you think whether this minimum amount of supervised learning affects the definition of "unsupervised" for MCG, or DeepCO3?

Thanks.

Hi, Great work! It is very promising.

I have some problems about the training procedures and datasets:
(1) The training of co-peak module, and the testing of final instance co-segmentation share the same datasets, i.e. COCO-VOC, COCO-NONVOC, VOC12, and SOC? I am confused about this.
(2) The training of co-peak module is unsupervised or weakly supervised (images sharing the semantically related objects), thus the validation is conducted upon the three losses, is that right? Because you indicate that the optimization procedure stops after 40 epochs.
(3) The training and validation data of co-peak module are the same, but shuffling in each batch?

Thanks.

Hello,
the provided links to the google repositories are down.
Is it possible to reupload them?
I am in particular interested in the dataset.

Best regards

I've run the demo for 22 categories. However, the result seems that there is a difference compared to your result. And I don't know my experiment's problem.

Thanks a lot!

Hi, @KuangJuiHsu ,

You compare your method with C2SNet, a saliency detection method, I, therefore, tried to reimplement the instance segmentation result based on the contour map(from C2SNet) and the MCG method. However, I met the problem that the contour map generated by C2SNet is different from the contour map of UCMs, which finally leads the MCG cannot generate segment proposals.

Could you give me any suggestions about your implementation detail?

Thanks in advance.
Xin

Hi,

That's a very interesting work and I have a question about this paper and hope you can help me.

You resize the images to 448x448 and stimulate the co-peak with fixed 3x3x3x3 size. However the object sizes are various in different images, how does the co-peak loss to handle this varience?

Thanks.

Hi, could you share the results of all other instance-aware results compared in Table2 of your paper,
to promote the rapid development of instance co-segmentation, thanks a lot!