84 points by sportsvision00 1 year ago flag hide 26 comments
yanglab 1 year ago next
Great work! This is really impressive. I'm curious to know how you deal with occlusions when multiple players are in close proximity?
yanglab 1 year ago next
@codingbandit we used a combination of background subtraction and mask R-CNN to handle the preprocessing.
yanglab 1 year ago next
@mlwhiz That's an interesting idea, we will definitely look into it! Thanks for the suggestion.
johndoe 1 year ago next
@codingbandit Exactly! And the choice of model really depends on the specific use case and the resources available.
codingbandit 1 year ago next
In terms of real-time analysis, how do you handle sending and processing the data? Do you use any streaming technologies?
johndoe 1 year ago next
We use Kafka for streaming the data and processing it in real-time. It's been a great solution for us so far.
mlwhiz 1 year ago next
We've used Kafka as well, and it integrates nicely with Spark Streaming for large-scale real-time data processing.
codingbandit 1 year ago next
I'm curious about how you handle serving the models for real-time predictions. What technologies or frameworks do you use?
johndoe 1 year ago next
We've used TensorFlow Serving as well, and the API is quite intuitive. But we've also found that using microservices with gRPC can be a good alternative solution.
mlwhiz 1 year ago next
True, but it also requires more development and maintenance work. It's always about finding the right balance based on the specific needs and resources.
mlwhiz 1 year ago prev next
I've also been working on similar problems, have you looked into using pose estimation to help with occlusions?
codingbandit 1 year ago next
@johndoe Yes, YOLO can be faster than other object detection models, but it can sacrifice some accuracy. It's all about trade-offs.
yanglab 1 year ago next
@mlwhiz That's a good point, we will definitely look into ensemble methods for future improvements. Thanks!
codingbandit 1 year ago prev next
Really cool project! I'm interested in the preprocessing steps you used to prepare the images for the ML model?
johndoe 1 year ago next
For this kind of application, have you considered using real-time object detection models like YOLO?
mlwhiz 1 year ago next
I've found that using ensemble methods can help improve the accuracy with real-time object detection models. It's definitely worth exploring!
johndoe 1 year ago next
Have you considered using transfer learning to speed up the training process for the ML model?
mlwhiz 1 year ago next
Yes, transfer learning is a great way to save time and resources. We've used it in our own projects with good results.
codingbandit 1 year ago next
@johndoe That's great to hear. Kafka is a reliable solution for real-time data streaming.
johndoe 1 year ago next
Yes, Spark Streaming is definitely worth checking out! We've used it in conjunction with Kafka and it's quite powerful.
mlwhiz 1 year ago next
We use TensorFlow Serving for serving our ML models, and it's been great so far. It's easy to deploy and manage the models.
codingbandit 1 year ago next
@johndoe That's a good point. Microservices with gRPC can offer more flexibility and customization for specific use cases.
sportstechie 1 year ago prev next
Very cool project! I'm curious if you have any plans to apply this to other sports or use cases?
ai_engineer 1 year ago prev next
Great work! I'm also interested in knowing the hardware and infrastructure you used to support real-time processing?
nerd_alert 1 year ago prev next
I'm a big fan of ML and sports! Have you considered using the models to generate real-time statistics or predictions during the game?
dataguy 1 year ago prev next
Cool use of computer vision and ML! Can you talk a bit more about the evaluation metrics and techniques you used to assess the model's performance?