Segment Anything 2 (SAM 2) Ball Tracking and Real Time Code Demo!

[@kevinwoodrobotics]

Learn Computer Vision using AI: kevinwoodrobotics.com/product/computer-vision-using-ai/ Code and Doc: kevinwoodrobotics.com/product/segment-anything-2-sam-2-ball-tracking-and-real-time-code-demo/

[@deeplearningexplained]

Solid tutorial, I really enjoyed the coding part! 👍

[@kevinwoodrobotics]

Thanks!

[@davidlennear3112]

Excellent video. I want to do the same thing but for static images. Like floor plans. Any suggestions?

[@kevinwoodrobotics]

Yeah it should work too. You just need to choose the seed appropriately and see how it performs

[@Innovan3]

I cant seem to find any minimum specs to run this model on the website, what hardware specs would you recommend to run this in real time at 30 or the full 44 frames per second?

[@kevinwoodrobotics]

the tiny model is about 2x faster than the large model which i showed in my video. For reference, I'm using a i7 with 16 gb of ram and a pretty weak gpu 1650 Ti. I think with a more powerful gpu like the 40xx series and the tiny model might be able to meet your performance needs, but as with anything best to test it out since you program can also influence the performance

[@adlihs]

A stupid question: If you use it with players you can get coordinates x,y of player location?

[@kevinwoodrobotics]

Yes that’s right

[@williamshen4808]

Is it possible to use SAM2 to find the speed and acceleration of an object if you know how far its travelling?

[@kevinwoodrobotics]

Yeah it should be. Would be an interesting application

[@shailendrasingh6154]

Can you please share a some steps to run demo locally in my APPLE M1 Pro , 16 GB memory ?

[@kevinwoodrobotics]

The general flow should be similar as described in my doc with some minor MAC specific commands

[@paulpolizzi3421]

This looks amazing, how long can you track the ball? will it track it for the whole half of the game? cant wait to see what else you can do? awesome!!!

[@kevinwoodrobotics]

Yeah it’s pretty awesome! I haven’t tested that yet but would be an interesting test

[@ashishreddy2634]

Hi Kevin, I have a trained YOLO detection model and I am trying to develop the tracking dataset for my project. Just curios to know that can we do only bboxes tracking using SAM2 ? Cause so far I am using CVAT for annotating the videos in track mode and it is a time taking process. Any suggestions would be helpful. Thank you.

[@kevinwoodrobotics]

You can do semantic segmentation with sam2

[@eltounytactics1585]

Can I use this code to also track the players?

[@kevinwoodrobotics]

Yeah should be

[@BrunoEsteves-j9l]

Would it be possible to use it in a use case where I want to identify in a poker game not only where the cards are but what specific cards are on the table?

[@kevinwoodrobotics]

I think yolo might be better for this, but would have to try. I think for some cards where they look different enough it should work with SAM2 but if they are too similar it may get confused like the example with the horses

[@kumardeepam]

Hi Kevin, I need help getting the code to work. Whats the best way to reach you?

[@kevinwoodrobotics]

Feel free to email me or message me over LinkedIn. Info in my bio

[@deemon101]

Ok, but it seems like the difference between yolo and SAM is that with yolo, I can train it to detect a specific object and it doesn't matter which video or image I feed to it, it will keep detecting that object that I trained it to detect. Can this be done with SAM 2 too?

[@kevinwoodrobotics]

That’s a good point! You’re right for general detection yolo will still be needed. But for a single video SAM2 would work. The two can probably be combined to help with tracking

[@deemon101]

@@kevinwoodrobotics yeah, I guess this has already been done with the first SAM model here: kzbin.info/www/bejne/j5W6qJaVm9xjr7Msi=g5cgk3EJbHIIbqIG But then, I'm not sure why would I use both SAM and Yolo if yolo technically is doing the same thing just.. you have to train it first. I guess in the end SAM 2 is simply for different use case. Edit: ok maybe it's useful for faster segmentation annotation jobs as you can just draw boxes around the things you want to segment in yolo, then combine yolo with SAM 2, and magic happens... hmm

[@kevinwoodrobotics]

@@deemon101 yeah besides the speed you get with sam2, the other main benefit is yolo by itself sometimes struggle with tracking when there’s occlusion which sam2 solves

[@danialothman]

Hi Kevin, what is your PC specs?

[@kevinwoodrobotics]

I have an i7 with 16 gb of ram using a 1650 ti

[@kumardeepam]

than you for this video.

[@lillol3245]

really? money for the code?

[@kevinwoodrobotics]

Yes

[@KarmikBarodia-co7sq]

@@kevinwoodrobotics why can't you provide the code