zimmerrol / mask-rcnn-edge-agreement-loss Goto Github PK

First of all, your work is definitely of great significance! I want to implement your method in my project: a cascaded mask rcnn. But I'm wondering whether to use the edge loss. Is it better to use the loss three times or simply use it in the last stage of cascaded mask rcnn? Do you have suggestions for that? I'm really appreciated your reply!

Hi, thanks for your nice work and I appreciate it much.
I have some questions about your implementation. Your implementation is based on Mask_RCNN but I found this implementation might exist many problems which led to much lower performance than the official implementation(Detectron, maskrcnn-benchmark).
I'm interested in your work and try to implement your proposed method in maskrcnn-benchmark. The only difference between my implementation and yours is that I use abs() instead of sqrt() to aggregate edges detected from X-direction and Y-direction because sqrt will result in numerical problems. sqrt and abs is nearly the same in the theory.
And I obtained results below and my implementation is consistent with the official implementation.

And I wonder why I can't obtain the performance gain in your paper. Can you provide results obtained from more accurate implementations?

Thank you very much for the useful implementation.

When I try to load my last weight file with

!python3 top.py train --dataset=/content/drive/'My Drive'/Mask_RCNN/samples/top/dataset --weights=last

It generated another log folder and gave me configuration errors.

I have notice that you added
if self.config.RUN_NAME: key += "_" + self.config.RUN_NAME
to model.find_last() different from matterport. May I ask you what is the purpose of it?

To fixed my error, I had to remove the newly added code and add dir_name = os.path.join(self.model_dir, dir_names[-2]) as here

Detailed answers would be very helpful.
Thank you in advance

The code need the correct weight to calculate the total loss. By adding the weight in the LOSS_WEIGHTS

"mrcnn_mask_edge_loss":1.

And by the way
These line are missing in the config file too

    EDGE_LOSS_SMOOTHING_PREDICTIONS= False
    EDGE_LOSS_SMOOTHING_GT= False  

You code have result Low accuracy boundaries like this

How to resolved for high boundaries ?

Hi @JulienSiems ,

Thank you very much for a great contribution. I am working to family with Mask-rcnn.
Could you guide me how to run the code. Do I need to start training again ?
Everything are same with original Mask RCNN right?. Did you only add some functions on model.py ?
Thank you very much.

Hi,

I was trying to run mask rcnn edge agreement on my own dataset and tried to segment with this but results are not as expected. I also ran the original Mask rcnn on these images and results are ok except at edges and tricky corners. I train the network with edge agreement as python3 liver_CT.py train --dataset=/Mask_RCNN/mask-rcnn-edge-agreement-loss-master/dataset/Liver_CT_YIQ/ --weights=imagenet. Same as the original mask rcnn expect using new codes.

Can you help me to know whether I have trained the data correctly or I need to do some more changes for training on your own data.

Thanks,
Supriti

Hi I am working on a project that will detect cracks using mask rcnn. I wan to implement your code however I already have done annotating my 3,500 images on VGG annotator. I have my json file already. As I looked into your code, I see you have a .py file for annotation tool. In order to train my model with your code, do I need to annotate again? How can I use my .json file to your code? Hoping for a reply. Thank you