Missing bbox and frames about vibe HOT 11 CLOSED

@athn-nik I just tried it again on GoogleColab, now its working fine with Pytorch 1.4. By the way when training code will be released?

from vibe.

Hi @lan786,

Could you put a breakpoint somewhere after this line, and check if multi object tracker is working?

from vibe.

@mkocabas Thanks for your response.

I just gave a try as you suggested in above comment. tracking_results on this line showing empty dict. I'm sure about it that there should be bbox and frames in tracking_results.

Furthermore at line 71 in num_frames there are 301 frames, which seems that frames are there but MOT is not picking up frames and unable to define bbox.

If I use maskrcnn as detector then all the things are fine and project is working but as you compare Runtime Performance maskrcnn is slow and need higher memory. In my case (RTX2060) tracker-batch-size with 12 is out of memory. I can run it maximum batch size of 8, but in case of Yolo it seems that batch size with 12 is also fine with RTX2060.

How I can overcome this problem to use YOLO as detector?

from vibe.

Could you try running this standalone: https://github.com/mkocabas/yolov3-pytorch? This is the package I use for YOLO.

from vibe.

Hi @mkocabas I just gave a try to run your updated yolov3-pytroch as standalone and I got some issues I submitted new issue in that repository, that issue is also mentioned above. Kindly check that and suggest me some valuable solutions. I want to use yolo as detector in VIBE.

Furthermore, using maskrcnn as detector with torch 1.4, I got errors. I realized that this error is due to pytorch so I changed pytorch from 1.4.0 to 1.2.0 and it works fine. The same problem arise in GoogleColab with pytorch 1.4.0 and maskrcnn, but using yolo as detector pytroch 1.4.0 also works fine in GoogleColab.

from vibe.

We used pytorch1.4 and it works fine. Have you both ways (cond and virtualenv) and have the same issues? Could you post your error log here and then I could give it a try to solve it. @lan786

from vibe.

I keep seeing people ask about training code.....if this code doesn't include the training code then how is it possible to run it?

from vibe.

The code is expected to be released by mid April.
@Primate92587 as every ml model during inference. You have a trained model, you load the weights and network(here instantiate the network class) and then you run the model by giving it the data that it should be fed(here videos).
Then, the model gives you the output and you present it somehow(here with a video).

from vibe.

@athn-nik yeah i know how ml models work...just wanted to verify that the code can't be run right now because it's basically incomplete?

from vibe.

Ah, ok. No, the code runs just fine. It is a demo code. A code that shows our results using an already trained model. People are just asking for training code to train a model from scratch, experiment with it and see exactly how everything works.

from vibe.

Got it. thanks.

from vibe.