6 ms·

The Toronto Raptors, SportVU cameras, and the NBA's analytical revolution

112 points by swanson 14y ago

mmc 14y ago
The "ghost" players in these animated replays that show where the system thinks a player should have been are fascinating to see. I wonder how automatic it is - if games could be analyzed like that without much expert input, this would be a really cool thing to have for recreational teams who want to get better but can't afford expert coaching. I'm thinking my late-night adult league hockey team would probably benefit from some replays like this.
- swanson 14y ago
  It would be interesting - though it will be some time before the price comes down. The cameras to do the tracking cost $100k/year at the moment.
  - mmc 14y ago
    $100k/year is certainly a lot :) 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.
  - eumenides1 14y ago
    While I agree 100K is alot, it's nothing compared to salary and cap space. To be able to spend money outside of cap restictions to imporove your team is a stragetic advantage. 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.
- streptomycin 14y ago
  I'd be pretty curious about that too. I actually tried doing something semi-similar before. My goal was to simulate a basketball game by having every player take an action based on "expected points per possession". So if you are covered tightly but a shooter is open, your expected points would be high for passing but low for shooting. This would then be done recursively for every decision, continuously looking ahead until the shot clock expires (so maybe the best move would be to pass it into the post, draw a double team, and then kick it out for an open shot). 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?
  - abecedarius 14y ago
    http://en.wikipedia.org/wiki/Markov_decision_process http://en.wikipedia.org/wiki/Markov_decision_process for a start.
- the_watcher 14y ago
  Unfortunately the cameras are pricey enough that only 15 of 30 NBA teams have purchased them, not to mention the half decade of work it took the Raptors to build the code to actually generate the visualizations. Maybe someday it will be commercially available and affordable, but that day seems a long way off to me.
  - nickalewis 14y ago
    I don't think it's the cost that's keeping other NBA Teams from purchasing ($100K/year is pennies to an NBA franchise). I suspect its more of a front-office vs coaching thing. Adopting this as part of a teams core prep takes education; Effectively learning how to understand this type of data and apply it in practice over the course of a season.
    - the_watcher 14y ago
      It should be pennies to an NBA franchise (when they are spending millions on players who don't play). Unfortunately, owners like Robert Sarver and Donald Sterling are notorious for refusing to spend on marginal upgrades like a quality visitors locker room, or a a full support staff. Even Paul Allen has a reputation for refusing to pay for a high quality training staff in Portland - probably a bit undeserved since he seems to spend on other important things and I can't imagine why training staff would be something he ignores. Also, there is a floor for player salaries that must be met. The analytics vs. coaching issue is a real one too, but don't discount owners being cheap.
  - ssharp 14y ago
    I'm wondering why the 15 other teams aren't on board yet. If the price is only 100k, it's nothing to an NBA team. If those other 15 teams don't already have an advanced analytics team in place, then it makes sense not to have the camera until you have a structure in place to make sure of the camera data. 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.
- JoelSutherland 14y ago
  It sounds like there probably wouldn't be enough data to generate the ghost players. According to the article they depend on the skills of the individual covers. (i.e. Player X is a 35% three point shooter so don't leave him in the corner to help on a cut) 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.
balakk 14y ago
The camera presumably is overhead. From that perspective, what could be a useful feature for person tracking? I am assuming it wont be that easy to pickup just numbers of their jerseys. Faces are also not easy to get from that angle. Or do they have multiple cameras in place?
- thespons 14y ago
  I could be wrong, but I would guess that they're not mounted overhead for that reason. At least not directly overhead. Since the court is a known, fixed size, I think you'd be able to project the players' location from a normal camera view onto an overhead representation like they show in the article's videos.
- pitt1980 14y ago
  probably use 2 - 3 camera to triangluate location if you only have to idenitify 24 people, you could probably do that on height plus some other physical dimension
  - osdf 14y ago
    Or with some Machine Learning: Identifying Players in Broadcast Sports Videos using Conditional Random Fields, CVPR 2011.
    - pitt1980 14y ago
      cool link, thanks 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.” http://www.thesmartset.com/article/article01181302.aspx http://www.thesmartset.com/article/article01181302.aspx
nickalewis 14y ago
Great article. As a huge NBA & NFL fan, It's intriguing to see the front offices of some of these teams begin to think outside of the box when it comes to player development. I think that's the most interesting opportunity here -- the practice of "watching film" is now innovative enough to not only help prepare for opponents, but also provide coaches a better macro view of what to work on with specific players in the short & long term. Especially in coaching behavior "off the ball". Thats always a game changer. 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!
- osdf 14y ago
  Btw. Shane Battier (back then with the Rockets) was mentioned in a new york times piece on NBA and Stats some years ago: http://www.nytimes.com/2009/02/15/magazine/15Battier-t.html?pagewanted=all&_r=0 http://www.nytimes.com/2009/02/15/magazine/15Battier-t.html?...
kevinthew 14y ago
To the uninitiated, and those who like sports analytics (specifically the NBA), you should absolutely follow Zach Lowe's articles on Grantland and elsewhere. He's the best sports writer on the subject, and probably the only coherent one at that.
swang 14y ago
All the analytics in the world won't save you if your GM is Colangelo.
wazoox 14y ago
Wow, memories. In 1998 I worked on the FIFA World Cup graphics, and two guys on the team were developing a camera player tracking and stat acquisition system. Back then, raw CPU power was a real limitation; the fastest PCs were PII 333 Mhz and they just couldn't track automatically 22 players from 4 cameras (one atop each corner of the stadium). 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 :)
- GFischer 14y ago
  Some former classmates of mine are still working on it, they provide analytics for the Uruguay soccer team among others: http://www.kizanaro.com/web/index.php?lang=en 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.aspx?trid=1334 http://www.technologyreview.es/tr35argentinayuruguay/profile...
recroad 14y ago
As a long-time Raptors fan and the co-founder/owner of the most read Raptors blog, I can confirm that none of this technology has helped the Raptors.
- muyuu 14y ago
  Seems to me like too many players in the Raptors are lacking stamina for such an intense defense system as that system produces. They would be gassed in no time.
awongh 14y ago
that is some really cool tech. I wonder if the camera reads the player's numbers on their jerseys when they go onto the court? 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.

112 points by swanson 14y ago

mmc 14y ago

The "ghost" players in these animated replays that show where the system thinks a player should have been are fascinating to see. I wonder how automatic it is - if games could be analyzed like that without much expert input, this would be a really cool thing to have for recreational teams who want to get better but can't afford expert coaching. I'm thinking my late-night adult league hockey team would probably benefit from some replays like this.

swanson 14y ago
It would be interesting - though it will be some time before the price comes down. The cameras to do the tracking cost $100k/year at the moment.
- mmc 14y ago
  $100k/year is certainly a lot :) 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.
- eumenides1 14y ago
  While I agree 100K is alot, it's nothing compared to salary and cap space. To be able to spend money outside of cap restictions to imporove your team is a stragetic advantage. 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.
streptomycin 14y ago
I'd be pretty curious about that too. I actually tried doing something semi-similar before. My goal was to simulate a basketball game by having every player take an action based on "expected points per possession". So if you are covered tightly but a shooter is open, your expected points would be high for passing but low for shooting. This would then be done recursively for every decision, continuously looking ahead until the shot clock expires (so maybe the best move would be to pass it into the post, draw a double team, and then kick it out for an open shot). 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?
- abecedarius 14y ago
  http://en.wikipedia.org/wiki/Markov_decision_process http://en.wikipedia.org/wiki/Markov_decision_process for a start.
the_watcher 14y ago
Unfortunately the cameras are pricey enough that only 15 of 30 NBA teams have purchased them, not to mention the half decade of work it took the Raptors to build the code to actually generate the visualizations. Maybe someday it will be commercially available and affordable, but that day seems a long way off to me.
- nickalewis 14y ago
  I don't think it's the cost that's keeping other NBA Teams from purchasing ($100K/year is pennies to an NBA franchise). I suspect its more of a front-office vs coaching thing. Adopting this as part of a teams core prep takes education; Effectively learning how to understand this type of data and apply it in practice over the course of a season.
  - the_watcher 14y ago
    It should be pennies to an NBA franchise (when they are spending millions on players who don't play). Unfortunately, owners like Robert Sarver and Donald Sterling are notorious for refusing to spend on marginal upgrades like a quality visitors locker room, or a a full support staff. Even Paul Allen has a reputation for refusing to pay for a high quality training staff in Portland - probably a bit undeserved since he seems to spend on other important things and I can't imagine why training staff would be something he ignores. Also, there is a floor for player salaries that must be met. The analytics vs. coaching issue is a real one too, but don't discount owners being cheap.
- ssharp 14y ago
  I'm wondering why the 15 other teams aren't on board yet. If the price is only 100k, it's nothing to an NBA team. If those other 15 teams don't already have an advanced analytics team in place, then it makes sense not to have the camera until you have a structure in place to make sure of the camera data. 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.
JoelSutherland 14y ago
It sounds like there probably wouldn't be enough data to generate the ghost players. According to the article they depend on the skills of the individual covers. (i.e. Player X is a 35% three point shooter so don't leave him in the corner to help on a cut) 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.

balakk 14y ago

The camera presumably is overhead. From that perspective, what could be a useful feature for person tracking? I am assuming it wont be that easy to pickup just numbers of their jerseys. Faces are also not easy to get from that angle. Or do they have multiple cameras in place?

thespons 14y ago
I could be wrong, but I would guess that they're not mounted overhead for that reason. At least not directly overhead. Since the court is a known, fixed size, I think you'd be able to project the players' location from a normal camera view onto an overhead representation like they show in the article's videos.
pitt1980 14y ago
probably use 2 - 3 camera to triangluate location if you only have to idenitify 24 people, you could probably do that on height plus some other physical dimension
- osdf 14y ago
  Or with some Machine Learning: Identifying Players in Broadcast Sports Videos using Conditional Random Fields, CVPR 2011.
  - pitt1980 14y ago
    cool link, thanks 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.” http://www.thesmartset.com/article/article01181302.aspx http://www.thesmartset.com/article/article01181302.aspx

nickalewis 14y ago

Great article. As a huge NBA & NFL fan, It's intriguing to see the front offices of some of these teams begin to think outside of the box when it comes to player development. I think that's the most interesting opportunity here -- the practice of "watching film" is now innovative enough to not only help prepare for opponents, but also provide coaches a better macro view of what to work on with specific players in the short & long term. Especially in coaching behavior "off the ball". Thats always a game changer. 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!

osdf 14y ago
Btw. Shane Battier (back then with the Rockets) was mentioned in a new york times piece on NBA and Stats some years ago: http://www.nytimes.com/2009/02/15/magazine/15Battier-t.html?pagewanted=all&_r=0 http://www.nytimes.com/2009/02/15/magazine/15Battier-t.html?...

kevinthew 14y ago

To the uninitiated, and those who like sports analytics (specifically the NBA), you should absolutely follow Zach Lowe's articles on Grantland and elsewhere. He's the best sports writer on the subject, and probably the only coherent one at that.

swang 14y ago

All the analytics in the world won't save you if your GM is Colangelo.

wazoox 14y ago

Wow, memories. In 1998 I worked on the FIFA World Cup graphics, and two guys on the team were developing a camera player tracking and stat acquisition system. Back then, raw CPU power was a real limitation; the fastest PCs were PII 333 Mhz and they just couldn't track automatically 22 players from 4 cameras (one atop each corner of the stadium). 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 :)

GFischer 14y ago
Some former classmates of mine are still working on it, they provide analytics for the Uruguay soccer team among others: http://www.kizanaro.com/web/index.php?lang=en 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.aspx?trid=1334 http://www.technologyreview.es/tr35argentinayuruguay/profile...

recroad 14y ago

As a long-time Raptors fan and the co-founder/owner of the most read Raptors blog, I can confirm that none of this technology has helped the Raptors.

muyuu 14y ago
Seems to me like too many players in the Raptors are lacking stamina for such an intense defense system as that system produces. They would be gassed in no time.

awongh 14y ago

that is some really cool tech. I wonder if the camera reads the player's numbers on their jerseys when they go onto the court? 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.