search

gaomingqi / Track-Anything

Track-Anything is a flexible and interactive tool for video object tracking and segmentation, based on Segment Anything, XMem, and E2FGVI.
MIT License
6.44k stars 480 forks source link

Still have not had a response- Requesting help ASAP #107

Open Artisto1 opened 1 year ago

Artisto1 commented 1 year ago

I have posted a previous request. I am a video editor and compositor desperately trying to open doors for linux use. There is not a clear direction for a still new linux user to be able to install this on a particular distro. At the moment, I am experimenting with MINT. Though I copied the command lines and think I installed the dependencies- How do I run the program? I don't seed it in my installed programs after searching by name. It has been some time now and I am asking for some immediate clarification so that we can all break away from any need to ever use Windows again!

jin-eld commented 1 year ago

@Artisto1 you are dealing with cutting edge stuff which is in development, don't expect a Windows-like installer and an icon to launch the app; as much progress as Linux has made in graphical interfaces imho you need to learn at least some basics on the command line if you really want to switch and utilize it's full potential. Especially when you want to use all those AI things which are under heavy development - you won't get far without the command line.

That being said, the installation instructions mention all steps: https://github.com/gaomingqi/Track-Anything#linux--windows

The last step launches the application:

# Run the Track-Anything gradio demo.
python app.py --device cuda:0

Provided that you followed all the required installation steps and downloaded the models, the application will start and you will see a URL in the console - that's where you navigate to in your web browser in order to get to the Track-Anything user interface.

There are some issues that I think the authors did not yet address, for instance the --device parameter is ignored and is always hardcoded to cuda:0.

As far as using Track-Anything for production: in my tests the masking is not yet good enough, depending on your video. It will easily produce errors (i.e. extend onto other objects) which need to be corrected via negative masks or manually.

I failed to use longer videos with Track-Anything (and by "longer" I mean only 3 minutes) on a 16GB GPU - I ran into a GPU OOM. The video worked on the CPU on a 64GB system, but of course took a lot longer to render.

So while this is very exciting and a starting point for the future, don't expect any miracles just yet.

P.S. I am also just a user and not related to the project, so what I wrote are just my 2 cents.