yihongxu / deepmot Goto Github PK

Is 'Car/Vehicle' one of your training dataset as well? I am using MOT17 dataset for my research as well but I am not sure if it includes Cars/Vehicles.

Are these clean_detections filtered by a confidence threshold or got by anthor algorithm？

Hi, thanks for uploading the training code for DHN!
I have some questions for DHN:

1. How did you create the train and test split? On the paper, it said that you had 114,483 and 17,880 for training and testing splits respectively, but when I ran your code, the validation and train lengths for 1 batch size were 1998 and 11870.
2. While training, I saw precision and recall were steadily increasing, but weighted accuracies were 0.00% throughout. Could there be a bug in your code?

I need to train DeepMOT on my custom dataset. Is it necessary to train Deep Affinity Network on the custom dataset?. For training the Deep Hungarian network is there code available?

This is the result of running _results.py

Thank you for your sharing. However, in tracking_on_mot.py, your implementation of siameseRPN is not parallel, which lead to very slow inference time. Could you provide a parallel version. Thank you.

How to run pretrained network on video ?

Thanks for sharing. I want to test the pre-trained model on my own data. How could I change the file path of the testing data?

I need to use the DHN in my project. I want to ask how to use this method to compute identity switches? Can you add more detailed explanations for each of the parameters?

Hi，thanks for your sharing! can you give some explanation about the code of the birth and death process? the idea in paper you said is simple but the code looks so complex that I can not understand. please tell me your coding idea in a simple way,like fisrt you want to fiter bboxes with IOU,second use appreance feature....?

I want to apply your great work to my private data.
I have run your code on MOT successfully. But I met some difficulties when generating DHN_data on my private data.
I have read your paper but am still confused.
Could you explain how to generate the DHN_data on MOT in detail and provide a README for the DHN_data?
Could you please share the code for generating the DHN_data?
Thanks! @yihongXU

I download your pretrained model and followed your instructions to test in MOT16 training dataset. The results I got like this:

Have you test your model in other dataset?
Is the pretrained model the newest one?

What is its fps? Can it be real-time?

Hi, Xu! Thanks for sharing your work.

Could you please offer the version of the python, pytroch and torchvison in this project?
And also the verisons in Tracktor（from phil bergmann） : torch=0.3.1, torchvion=0.2.0. Is that compatible with deepmot?

Thanks

Hi,

How did you go about training DHN? From my understanding of the paper, you trained it as a standalone module (by itself, without the tracker and differentiable MOTA and MOTP, is that correct?)

Also, according to the paper, you used the focal loss as the supervisory signal for training the DHN. How did you compute for focal loss?

Thank you.

when i test the pretrained model, i meet this mistake.

Hi, thanks for your great work!
One simple question about DHN:

You claim that DHN enables end-to-end training of deep multi-object trackers, but DHN is pretrained and fixed during training.

My question is: why not use hungarian matching? In my opinion, it always gives you the exactly right matching results. And I cannot find the comparisons of hungarian matching with DHN (from the aspects of speed/performance) in your ablations.
I've also observed that DHN (with two lstms) is much slower than Hungarian matching that runs on CPU in my project, so the speed issue may not be the reason I think.

Please correct me if I am making mistakes here! :)

I'd like to run your code on PyTorch 1.4.
But, the code in your master branch is only for PyTorch 0.4.1.
Is it possible to reproduce your results with obsolete branch?

The requirements.txt says that I should install Pytorch 1.3. But when I run test_tracktor/experiments/scripts/tst_tracktor_private.py, it said ImportError: torch.utils.ffi is deprecated. Please use cpp extensions instead., I tried to downgrade pytorch to 0.4.1 after asking on stackoverflow for help. But more errors raised!
I've commented out from torch.utils.ffi import _wrap_function in test_tracktor/src/frcnn/frcnn/nms/_ext/nms/__init__.py,and it raised a new error: ImportError: test_tracktor/src/frcnn/frcnn/nms/_ext/nms/_nms.so: undefined symbol: state.
I think it seem to be someting wrong on _nms.so, but I have no more solutions to fix this issue.

Best regards!

Why not upload the Tracktor version to verify the state-of-the-art?

Hi Yihong Xu. I would like it to be updated. I would like you to make a note on how it will work. good work

Hi,

May I know how you came up with the MOT17 test set metric results you provided in the readme? It seems that your evaluation.py code requires gt.txt files to generate these results. However, the gt.txt files are unavailable for the MOT17 test set. Thank you.

First, thanks for your great jobs and share.

I want to use your code in my research and want to restrict coordinates where new IDs are generated. (Currently, the new ID may be an implementation that can occur at any position of detections. )

If possible, please tell me how to implement it.

I meet the same issue as issue #40. The trained model works well according to the evaluation results. However, the training loss does not decrease during training just like issue #40

Thank you very much for sharing your code. I am very interesting for this paper and i have some improvements to this model, i hope sincerely that you can share the training code of Deep Hungarian Network.

When I try to run the provided singularity image (tracker.sif) I get the following error:

ERROR : Unknown image format/type: tracker.sif
ABORT : Retval = 255

What can I do?

notice that you use the DHN in training，
but use DAN in evaluate，
what is the difference, and how the DHN influence the tracking result ?

I have tried tracking_on_mot.py and train_mot.py. They both work fine.
But when i want to run plot_results.py, it only showed "loading parameters..." and no output.

What should i do? I only modified the curr_path in .py file.

thanks a lot!

Can anyone please provide a google colab / jupyter notebook for implementation?

consider you have trained the SiamTPN + DHN, and why do't use the optimal assignment matrix to get the final result, but use the DAN that have not trained together?

user command $ "git clone [email protected]:yixu/deepmot.git"

following error has been encountered:

Cloning into 'deepmot'...
Permission denied (publickey).
fatal: Could not read from remote repository.

Please make sure you have the correct access rights
and the repository exists.

Hi.
Since DHN can replace Hungrain algorithm, why DHN is only used in the training stage? How does it perform in the testing stage?

Many thanks.

it seems that there is not any training code for DHN model in this project. you just optimize the sot instead of DHN.
optimizer = optim.Adam(filter(lambda p: p.requires_grad, sot_tracker.parameters()), lr=args.old_lr)

Hi,
Thanks for your work.
When I run the script $ python train_tracktor/experiments/scripts/train_tracktor_full.py. There is an AttributeError: 'FPN' object has no attribute 'reid_branch', originate from the line 339 in function step_full_reid of train_tracktor/src/tracktor/tracker.py. i.e.,
gt_real_features = self.obj_detect.reid_branch(gt_features)

I check the Class FPN(FPNResNet) in script train_tracktor/src/tracktor/fpn.py，there is no function defined as 'reid_branch'.

Looking forward to your response. Thanks a lot.

Hi,
I am following your great work recently.
I download the pre-trained DHN model weights and train a tracktor.
I checked the log in result_full and got oscillating loss as the following:

I m drunk lol 😁

if I want to run the deepmot on my own data, should i retrain the sameRPN network？

Thank your for your work. It is first one to use SOT for MOT. I use 1080ti to train. But It is too slow. It will cost one day for six epoch. Your paper say it will use 6 hours for 20 epoch in titanxp . Is there any change in code??

I am getting this error :

FATAL: container creation failed: mount /proc/self/fd/5->/usr/local/var/singularity/mnt/session/rootfs error: can't mount image /proc/self/fd/5: failed to mount squashfs filesystem: invalid argument

Kindly give an example for this statement :

singularity shell --nv --bind :_ tracker.sif

Hi！Thanks for your decent work. I have a problem when I run tracking_on_mot.py as follow.Could you tell me what can I do for this problem.

Hi,

Can we perform detection and tracking of cars/vehicles as well by using this model? These objects are included in the MOT17 dataset is what I have learnt. Please correct me if I am wrong. If they are included, have those object classes been included in the training as well? I am not able to figure out this by going through the train_mot.py code. Thanks for your help in advance.

Hi, thank you very much for your contribution! Could you please share the code used to generate DHA_data? I'd like to reproduce your code, but I'm stuck here. I would be very grateful if you could!