Cases
Products
zDialog
OneDash
Vision Hub
ML-consulting
MAGE
TorchOK
Studios
Data Lab
Dialog Systems
About us
About company
AIC
News
Contacts
En
Ru
Portfolio
Products
zDialog
OneDash
Vision Hub
ML-consulting
MAGE
TorchOK
Studios
Data Lab
Dialog Systems
About us
About company
AIC
News
Contacts
En |
Ru
A neutral network that controls match broadcasts
Suggest a project
ABOUT
ABOUT
ABOUT
ABOUT
ABOUT
ABOUT
Main information
01/
Client
Task
Sportrecs.com - Russian startup, a sports broadcasting service.
Automate match streaming using AI-controlled cameras.
The neural network broadcasts games, monitors ball movement, and
switches between camera angles when necessary.
How it works:
Problems we have faced
02/
It is difficult for neural networks to recognize the ball because it is indistinct in the video.
01
There is not enough data to train the algorithm. To broadcast a match from a smartphone, you need a watch of mobile video
Processing multiple video streams requires a lot of computing resources
02
03
Development process
03/
Found source
Finalized
Added functions
Increased reliability
Merged with service
Released
Took a ready neural network that can track a big ball
Taught neural network to «see» big and small balls
Added model for recognition of players in the frame
Integrated neural network with Sportrecs service
Prepared a system for working with four synchronous real time video streams
How does the service work
04/
Smartphones are installed at different points on the field
Sportrecs streaming service starts on devices
The program connects to smartphones cameras
Every few seconds the neural network switches the camera so that the viewer does not get bored
Selects optimal plans where the ball is clearly visible
Demo video
05/
What have we learned
06/
01
02
Process multiple video streams simultaneously
Make the neural network fast enough for live mode
Demo video
Project team
07/
Computer Vision engineer
Project manager
Backend developer
MLOps
Techlead
Vladislav Vinogradov
Andrey Chertkov
Konstantin Kubrak
Aelita Shaikhutdinova
Alexey Zhuchkov
1 second of video is 30 frames, which the neural network must process
instantly
.
Interesting fact:
Our plans
08/
To teach a neural network to broadcast in any conditions, including large stadiums. Now it knows how to manage the show of games in the hall.
New conditions
Teach the neural network to control video cameras on tripods and cut the best moments of the match.
New features
New sports
Add new sports to the system, such as basketball, volleyball, hockey, boxing
Emotions from the project
09/
Konstantin Kubrak
Project Manager (Smashed a window with a ball during tests)
«99.9% of the world's population do not know about neural networks what we know now».
«While learning how neural networks work we also get to know ourselves»
{"0":{"lid":"1531306540094","ls":"10","loff":"","li_type":"nm","li_name":"name","li_title":"My name is","li_ph":"Michael","li_req":"y","li_nm":"name"},"1":{"lid":"1531306243545","ls":"20","loff":"","li_type":"em","li_name":"email","li_title":"My E-mail","li_ph":"mail@example.com","li_req":"y","li_nm":"email"},"2":{"lid":"1608756742134","ls":"30","loff":"","li_type":"ta","li_name":"Description","li_title":"Project description","li_ph":"Click here and describe your task in a free form","li_req":"y","li_rows":"2","li_nm":"Description"}}
Contact us