Hello! Just posting a first issue here after seeing this on the Discord.
I'm a open source machine learning researcher at the University of Florida with a decade of experience in the computer vision for airborne monitoring space, mostly with ecological objects, things like trees, birds, and whales.
https://deepforest.readthedocs.io/en/stable/getting_started/index.html#getting-started
I'm also an avid soccer fan and have plenty of access to compute and tokens. I'm learning how to get involved with this project and what you see as the major limitations here.
I'm volunteering my time for a local elite youth soccer academy and would like to revamp their clip generation process. They have access to Veo, a commercial cloud platform, but I was really unimpressed with its ability to autogenerate clips for anything besides goals? How much have you looked that commercial tools in this space, compared to modern computer vision, it seems very behind?
In terms of open source can you give me a sense of the state of the field, I saw your model weights on huggingface and a few open datasets?
A quick list of things a club might need at the small scale that I could build and contribute here:
- Clip generation for specific events for a given team
- Clip generation for specific players
- Simple metrics to connect with player individual development plans
Starting with simple things like: given these sets of videos, can you make 15-second clips of every goal kick for the team in blue? I've been able to create a fairly reasonable first attempt by scanning an image, passing to claude for visual understanding, resampling based on probability.
In general do researchers in this community publish model weights? Feels like this repo could be a good place to collate essential tasks, I see past competitions like https://github.com/SoccerNet/sn-grounding.
Give me a push on where to get started. I can train models, help with open source packaging, act as dev tester. Open to your ideas.
Hello! Just posting a first issue here after seeing this on the Discord.
I'm a open source machine learning researcher at the University of Florida with a decade of experience in the computer vision for airborne monitoring space, mostly with ecological objects, things like trees, birds, and whales.
https://deepforest.readthedocs.io/en/stable/getting_started/index.html#getting-started
I'm also an avid soccer fan and have plenty of access to compute and tokens. I'm learning how to get involved with this project and what you see as the major limitations here.
I'm volunteering my time for a local elite youth soccer academy and would like to revamp their clip generation process. They have access to Veo, a commercial cloud platform, but I was really unimpressed with its ability to autogenerate clips for anything besides goals? How much have you looked that commercial tools in this space, compared to modern computer vision, it seems very behind?
In terms of open source can you give me a sense of the state of the field, I saw your model weights on huggingface and a few open datasets?
A quick list of things a club might need at the small scale that I could build and contribute here:
Starting with simple things like: given these sets of videos, can you make 15-second clips of every goal kick for the team in blue? I've been able to create a fairly reasonable first attempt by scanning an image, passing to claude for visual understanding, resampling based on probability.
In general do researchers in this community publish model weights? Feels like this repo could be a good place to collate essential tasks, I see past competitions like https://github.com/SoccerNet/sn-grounding.
Give me a push on where to get started. I can train models, help with open source packaging, act as dev tester. Open to your ideas.