The Toronto Raptors, SportVU cameras, and the NBA's analytical revolution
grantland.com
grantland.com
The ghost AI is pretty impressive as well. I was skeptical at first as to how authentic this type of "player AI" would be, considering the program makes assumptions based on historical numbers and any athlete/coach would certainly argue that its different in-game, however this says a ton:
> That's brutal, and it's not a coincidence that the only team that consistently mirrors the help defense of its ghosts is Miami, Rucker says. The Heat have three of the best wing defenders in the league in Shane Battier, LeBron James, and Dwyane Wade, and the latter two are among the NBA's most gifted pure athletes. James can mimic DeRozan's hyperactive ghost in a way no other player can, Rucker says. "LeBron basically messes up the system and the ghosts," Rucker says. "He does things that are just unsustainable for most players."
Not a huge Miami fan (Go BULLS!) but you can't deny their defensive effort.
Thanks for sharing!
In the end the system was almost working, however we needed two people watching the TV stream and clicking on players to identify them when the automated system lost track. However it provided already tremendously detailed stats, but unfortunately nobody really cared :)
http://www.kizanaro.com/web/index.php?lang=en
most of the input is still manual as you mention.
They were recognized by MIT Technology Review :) (in Spanish)
http://www.technologyreview.es/tr35argentinayuruguay/profile...
I can imagine facilities like hockey rinks buying the cameras and selling the service to the teams that play there, but not at that price. My impression of hockey rinks is that they're not terribly profitable. Most of the ones I've played in seem like borderline charities.
I don't think the price will come down, the only people who need it are wealthy and the wealthy can't afford to not have it.
The obvious problem was that there are a ton of free parameters involved and tons of branching. Every little action creates a new branch of possibilities with its own set of probability distributions. So even if (as they describe in the article) helping out more often decreases the likelihood of the current ball handler making a shot, how does that propagate through the rest of the possession?
The analytics vs. coaching issue is a real one too, but don't discount owners being cheap.
It will be interesting with this huge push in basketball over the past couple years to see where this type of analysis leads. The NBA model right now is pretty simple: get two or three superstars, align them with solid role players, and play in championships. It's hard to imagine analytics pushing a team like Houston to be able to beat a team like Miami or Oklahoma City in a seven-game series because the talent gap is so substantial. Houston can play a smarter game but the other teams still function better. What I could see happening, however, if a team like Oklahoma City, while slightly less talented than Miami, can implement insights from this analysis to raise their EV, it might overcome the slightly smaller talent gap. Even in that small sense, it makes more to spend a couple million a year on analytics than it does to pay an average NBA player several million. Since the NBA has both a salary cap and maximum contracts, it pushes the price of an average play up quite a bit, so better value may be had by raising the effectiveness of your current talent than trying to bring in new players.
When the system gets more optimized and easier to use, the role of a coach & GM could get diminished, and with it, their salaries. Of course they would do everything in their power to stop that.
As someone who is on the periphery of working in a professional sport in both analytics and player development, the simple answer is the fact that market forces are comparatively weak here. You have monopolies with a ton of inertia in doing things the same way, plus a lot of the decision-makers are luddites in nature who think the human element is the single most important factor.
Times are changing as the economics of sports tightens and becomes more efficient, but it's not like there are ways to disrupt the market openly. The Dallas Mavericks (Cuban) and the Houston Rockets (Morey) are the ones leading the charge in the NBA, just like the Oakland Athletics (Beane), Cleveland Indians (Shapiro), and Tampa Bay Rays (Friedman) led the charge in MLB.
The 82-game NBA schedule really helps to generate large enough sample sizes about players.
I could see it being potentially possible to just use average data, but if your rec league is anything like mine, talent ranges from a guy can credibly fake a three and then dunk after one dribble to guys that can barely get up and down the court.
if you only have to idenitify 24 people, you could probably do that on height plus some other physical dimension
found this:
Our natural gait, it turns out, defines us as humans. Not speaking broadly — that we’re only truly bipedal mammal on the earth, blah blah blah — but as individuals. Researchers are increasingly convinced that how we walk can identify us as unique individuals, much like a fingerprint or retina scan.
This sort of research has been underway since the late 1990s, picking up more urgency after the 2001 terrorist attacks on New York and Washington, and the London train attack in 2005. DARPA and Homeland Security, among others, are keenly interested in video analysis programs that can separate out and analyze an individual gait, then use this like fingerprints. DARPA has been sponsoring “human identification at a distance” studies since 2000, which often combines gait analysis with facial and gesture analysis. It’s a hot field right now, and has been heralded as a less invasive approach than retina scans or blood tests or fingerprinting. One study I read on the algorithms of walking put the appeal simply: “Advantages include the fact that it does not require subject cooperation or body contact, and the sensor may be located remotely.”
"Three cameras on each side of the court..."
This reminds me of a talk I heard a long time ago from Takeo Kanade at CMU, who was working on reconstructing 3D scenes from multiple cameras shooting different angles of the same action. One example was basketball, and he had a demo showing a play reconstructed as if the camera was on the ball. Pretty cool stuff. It needed lots of cameras, though, and I remember asking him if they had explored how well you could do with a minimum # of cameras. I remember his reply being essentially that they were less interested in bad results with few cameras :)
Here's a link to some surprisingly old press about his work being used for Super Bowl 35 in 2001: http://www.ri.cmu.edu/events/sb35/tksuperbowl.html They used 30 cameras.
I could imagine a cool realtime consumer version of this for all kinds of sports where you get an instant replay type of thing projected onto some kind of madden/video game type avatars: unpause and play any game from any instant in the game you're watching.