DeepBall: how a neural network detects a ball in video

DeepBall is a specialized neural network model for ball detection in sports videos, designed to process images of any size and create a confidence map of the ball's position. It is based on a fully convolutional architecture and uses the concept of hypercolumns, combining feature maps from different levels of the hierarchy to improve detection accuracy by taking into account more visual context. DeepBall was originally developed for soccer, but it can be adapted for volleyball, given the similar tasks of detecting and predicting the trajectory of the ball.

Sources and documentation

1. Original article by DeepBall:
   • Title: DeepBall: Deep Neural-Network Ball Detector
   • Authors: Jacek Komorowski, Grzegorz Kurzejamski, Grzegorz Sarwas
   • Date: February 19, 2019
   • Link: arXiv:1902.07304
   • Description: This article describes the DeepBall architecture, which uses a one-pass convolutional network to detect balls in long-range videos. The model processes entire frames rather than individual patches, which improves performance compared to traditional methods such as Hough Transform or separate patch processing.

2. Related Studies:

   • A Trajectory-Based Ball Detection and Tracking System with Applications to Shot-type Identification in Volleyball Videos (2012)
   • Description: The study uses trajectory information to detect balls in volleyball videos, applying a Kalman filter to generate candidate trajectories and analyze the physical characteristics of the ball's movement. This could be useful for adapting DeepBall to volleyball.
   • An Effective Volleyball Trajectory Estimation and Analysis Method With Embedded Graph Convolution
   • Description: A YOLOv4 variant is proposed with an embedded graph convolution to predict the trajectory of a volleyball. This research demonstrates how modern neural network approaches can be adapted to quickly move a ball and change its scale within a frame.
   • Tracking of Ball and Players in Beach Volleyball Videos_ (2014)
     
   • Description: The study describes methods for tracking the ball and players in beach volleyball using background subtraction and parabolic interpolation of trajectories. The importance of choosing camera angles to reduce occlusions is emphasized.

3. Related commercial and open-source projects:

   • Balltime - AI Platform for Volleyball Highlights & Analytics
   • Description: The Balltime platform uses AI to automatically create highlights and analyze volleyball trajectories, including serve speed. This is an example of the commercial application of technologies like DeepBall.
   • VolleyVision GitHub: shukkkur/VolleyVision
   • Description: The project uses RoboFlow and YOLOv7-tiny models to detect and track the ball and players in volleyball videos. Includes a dataset of 25 thousand images and a DaSiamRPN tracker, which can be used as a basis for integration with DeepBall.