Where Will They Go? Predicting Fine-Grained Adversarial Multi-Agent Motion using Conditional Variational Autoencoders

Panna Felsen, Patrick Lucey, and Sujoy Ganguly

​

Paper review by Michael A. Alcorn

1. Introduction

3. Basketball Tracking Dataset

• 95,002 12-second sequences (sampled at 25 Hz, downsampled to 5 Hz for training) from 1,247 NBA games in the 2015/2016 season
• Sequence ends at a shot, turnover, or foul (i.e., no changes of possession)
• Transition plays are very important in basketball!
• Record player identity for trajectories of 210 players with most playing time
• Other individuals just use “canonical position” (e.g., point/shooting guard)

4. Methods

Like (batter|pitcher)2vec

4. Methods

vector containing “future” trajectories of players

vector containing historical trajectories of players

vector containing historical trajectories of players

“identity” (team or players)

vector containing historical trajectories of players

4. Methods

• How do you order the trajectories given different lineups?
• Build a tree of “templates” using expectation maximization (from 10 games from 2014/2015)

vector containing historical trajectories of players

4. Methods

Is this desirable? What information is being lost?

vector containing historical trajectories of players

4. Methods

4. Methods

In supplement

5. Experiments

Why doesn’t knowing the team/players help?

5. Experiments

5. Experiments