Dataset Link Used in GameState Model Training

[Proma17]

I'm a university student currently working on a project involving game state classification, and I've been using the model you shared in this repo. To help with my research, I’m looking to replicate the training process and was wondering if you could share the link to the dataset you used to train the model. Much appreciation!

[masouduut94]

Yes proma
Sure I will share the dataset.

FYI, I found the videos from youtube channel that I mentioned in the readme.md and simply annotated the start and end of the game-state like serving, playing and not playing. then wrote a script that can get all of the frames and save it as a video clip in its directory.
I share them soon. just have to find the video clips. And you have to know that it takes time to find the best possible sets for the model, because you have to remove some clips that does not fit the categories.

[suyunzzz]

Yes proma Sure I will share the dataset.

FYI, I found the videos from youtube channel that I mentioned in the readme.md and simply annotated the start and end of the game-state like serving, playing and not playing. then wrote a script that can get all of the frames and save it as a video clip in its directory. I share them soon. just have to find the video clips. And you have to know that it takes time to find the best possible sets for the model, because you have to remove some clips that does not fit the categories.

thanks for your excellent job! as your description, could i use this game state detector to badminton games?

[tdymel]

Hey,
amazing project!

I wanted to ask if you posted the datasets already somewhere.
I was looking around and wanted to play around with this model and the datasets.

[masouduut94]

Hi Tommy It is always nice to see that developers like my volleyball analytics project. I had little time to develop this project, so finally I found some time this weekend to focus on applying some of my ideas on this project. About camera angle: we always need x, y, z of the ball or other objects like players. in this project, having the court points and using perspective transform we can estimate the location of the objects on the ground (like players) and estimate the location of the ball when it is around the players on the court (when they touch ball while they are on the ground, so we know the ball x, y as well). So in reply to your question, if you can estimate and map the location of the objects on the court, so the camera angle can be different, but pay attention that the point estimation can get inaccurate. The performance: Speaking about performance brings two factors: 1. inference speed: Right now on the main pipeline we have: - VideoMAE: Utilized for classifying each 1 second of the game. - Ball segmentation: Utilized to detect ball. - Action detection: Utilized for detecting volleyball specific actions like serve, set, spike, ....

…

--- Court Segmentation:is not used in the pipeline. it is more like a utility to find the court lines fast and save them in database, so we can then use those lines on our stats gathering. So practically testing the code, I can see that for real-time events it is lacking certain things: - we need a queuing mechanism that can keep video frames on RAM as long as rally is on, and when the rally is over it passes it to the frame by frame procedure (yolo object detection) - If we could gather a dataset that includes ball + players + referee (we may need his gestures to find out the scored teams, etc) + volleyball actions (spike, set, block, etc), then we would be able to have a unified model capable of detecting every object. so this way we can enhance the model inference speed because the image is only goes through one model instead of 2 - 4 models. There are other software engineering techniques that we can use to optimize the speed (like Trt model, using multiple GPUs, or we can use c++ inference and extensions). I have written a medium article on how we can optimize speed using c++ extensions here. https://blog.devgenius.io/pythons-vision-c-s-strength-the-art-of-delegating-computationally-demanding-challenges-to-c-00fc86f7fc6a 2. accuracy: Regarding accuracy, as you know there are multiple things to be considered: 1. YOLO and VIDEOMAE model size. The bigger backbone of the model gets, the more accurate the model gets but the inference time decreases. so I suggest the nano, or the small version of the yolo. 2. Training configurations can improve the accuracy. 3. A thorough and perfect dataset of all objects (each class has to get at least 2K - 3K instances in as various as images it can get). 4. For getting valid results, pay attention to the test dataset way more than the training dataset. In early stages of the project development, I found myself training multiple models, yet not knowing which one actually is better than the other one. because although I could see the metrics and measurements, but the problem is that since the test dataset were not actually the same, so I couldn't actually understanding which one to choose. I know that you sent me an email and asking me about datasets. I have the ball, and volleyball actions dataset and I can share. but for the game status dataset. I have to spend sometime finding the videos that I used for the annotations. I used CVAT locally to annotate the videos. All in all, I am very pleased to see that simply developing a hobby project brings developers with the same interests and obsessions towards me. It is a great feeling and I am so proud of this. I have a lot of plans for it, but most of the time I get short of time. but don't worry. stay tuned for the changes, pal If you wanted to message me directly, you can message with this number: +989381224749 (available on whatsapp and telegram) I hope to see you catching your dreams soon. Bests, Masoud

On Thu, Aug 21, 2025 at 4:19 PM Tom Dymel ***@***.***> wrote: *tdymel* left a comment (masouduut94/volleyball_analytics#11) <#11 (comment)> Hey, amazing project! I wanted to ask if you posted the datasets already somewhere. I was looking around and wanted to play around with this model and the datasets. — Reply to this email directly, view it on GitHub <#11 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AFB4FCDKCA6OEZKT3PCFOO33OW55LAVCNFSM6AAAAAB55EDTPKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTEMJQGQ3DOMZWGU> . You are receiving this because you commented.Message ID: ***@***.***>