Colab with --tracking --pose_track?

[rschmaelzle]

Hello,
thanks for a fantastic project - it was the most accurate & easiest of all posetracking projects that I could get to run!
I used the Colab notebook provided https://github.com/tugstugi/dl-colab-notebooks/blob/master/notebooks/AlphaPose.ipynb and it worked pretty much out of the box (I ran into issues with a local GPU though). However, this notebook installs from the pytorch-subbranch of the repository, which is older.
I tried to install the master, but failed. In particular, I would like to obtain tracking results about person identity so that I can track multiple persons. However, the options --tracking or --pose_track are not working (it complains that the module 'detector' is missing - that seems to be only present in master).
Does anyone have suggestions for how to get the master branch to work on colab? I tried to using the instructions, but didn't succeed (the old pytorch branch had a requirements-file -- the setup/build isn't working as smoothly).
Thank you very much
R

[Fang-Haoshu]

https://colab.research.google.com/drive/14Zgotr2_F0LfvcpRi03uQdMvUbLQSgok?usp=sharing
Can you try if it works?

[rschmaelzle]

Dear Fang/Fred Haoshu!,
thank you very much. I had to make minor changes to the original notebook, but got it to work! I am digging in deeper as I am still struggling with two issues:
I) the tracked people sometimes swap identities (this has been noted before and if I run without --viz_fast it works mostly - if people stay in position).
II) one thing I ran into is what seem to be memory limits. With the googleColab, I get 12 GB RAM, but if I run the code on youtube videos that are longer than 15 sec, it crashes. I am now trying to see if I can get it to run on a local GPU, or upgrade Colab. Do you have a tip on the memory that is needed?
Again many thanks Ralf.

[Fang-Haoshu]

--qsize 100 can lower the memory usage

[rschmaelzle]

thank you. that seemed to work and is a great suggestion I will keep in mind!

• I tried to find out what what were meaningful values for --qsize (e.g does using '50' lower memory even more?)
• I still struggle running this on videos longer than 30 sec, such as 5 minute clip (which I what I need), and starting with a minute or so, it gets prohibitively long. I don't quite get why it is so slow, since many lower-quality tools (e.g. posenet for Javascript) work in realtime, and although the --pose_track understandably takes longer, it shouldn't be so slow (meanwhile I am working on ColabPro)
• the --viz-fast option causes the bounding boxes of the identified persons to be way off
  If these issues are meaningful, it would help me a lot to hear suggestions (and if not, the issue can be closed since my basic question has been answered - Thank you!)

[Fang-Haoshu]

Hi, if you want to run faster, you can set the conf in detector/yolo_cfg.py to 0.4 or higher.

[Fang-Haoshu]

--qsize 50 indeed lower the memory.
"--viz-fast option causes the bounding boxes of the identified persons to be way off"
Can you show how does it look like? Thanks

[rschmaelzle]

thanks a lot

• setting the yolo_cfg.py to 0.4
• and using size 50 worked as well
• Here is a screenshot of the issue with --vis_fast , and here is a longer video.

[Fang-Haoshu]

Oh I see, it is indeed a bug. Let me fix it.

[Fang-Haoshu]

Should be fixed.

[qhdqhd]

I meet same problem in colab, the options --tracking or --pose_track are not working ,too? can you tell me why?

[qhdqhd]

@rschmaelzle

[qhdqhd]

thanks bro, @rschmaelzle

[qhdqhd]

well i know. just add --vis works fine,which may be a bug, can not tracking without visualization

[dekathomas]

Hello @rschmaelzle, how can you visualize the bounding box?