reproduction of results

[Dzp1007]

Hi, I have tried your code in MIA2019, using tra_train.csv as training dataset and tra_val.csv as validation dataset. But after testing, I get a not very good result.
Specifically, in many frames, the cells are totally ignored. Thus when validating on Fluo-N2DH-GOWT1 training video02, I get only 5 cells tracked, and it seems that these 5 cells' tracking results are also incomplete.
Could you help me with this problem?
Thanks a lot!!

[christianpayer]

Thank you for your interest in our work. Sorry that the code does not work as expected. It's been more than a year since I worked with this code, so I don't know exactly what could cause your problem. However, I will investigate in that by running the code on my computer. Maybe I committed wrong parameters to this github repository. I will come back to you later, when I have more insights.

[Dzp1007]

Thank you for your interest in our work. Sorry that the code does not work as expected. It's been more than a year since I worked with this code, so I don't know exactly what could cause your problem. However, I will investigate in that by running the code on my computer. Maybe I committed wrong parameters to this github repository. I will come back to you later, when I have more insights.

I find that you have updated the csv files and I'm going to try these again to check if there's some improvements.
And this is so kind of you to provide such detailed help, and I am looking forward to your reply. Thank you for your help!!!

[Dzp1007]

Thank you for your interest in our work. Sorry that the code does not work as expected. It's been more than a year since I worked with this code, so I don't know exactly what could cause your problem. However, I will investigate in that by running the code on my computer. Maybe I committed wrong parameters to this github repository. I will come back to you later, when I have more insights.

Hi, have you found out the problems?

[christianpayer]

I finally managed to find some time to do some debugging of our code. Some paths, etc. were wrong and I corrected them.

However, I was not able to reproduce the problems that you described. I remember that I also encountered similar problems, if I use the Pyro dataset by running the server_dataset_loop.py and setting refill_queue_factor > 0.0. If refill_queue_factor > 0.0, training appears to be faster (as the GPU will always receive training images), but exactly the same training examples could be seen too often by the network, and it could extremely overfit.

I committed some changes and additional instance_neighbor.json files, which could make the preprocessing on the CPU much faster. I hope this can help you, too.

You could also look at the images generated when the training scripts are doing inference. Look there for images named 'embeddings_cropped_1.mha'. These images are multi-channel volumes. I attached some example images of what you should expect there.

Could you send me some of your output images? Maybe if I see your output images, I will get better insights to your problem.