Isolating Football Player Movement by Eliminating Camera Motion

http://phdfootball.blogspot.com/2014/09/isolating-player-movement-by.html

by tantalor • 12 years ago

156 35 12 years ago

phdfootball.blogspot.com

Disappointed to see this is about "American Football", not "Football".

Seriously, I'm curious about the ever-stricter rules about contrasting uniforms in football outfits - enough that national sides can't play in their classical colors anymore, and club teams have four to six alternate uniform sets; I thought that was to facilitate the automated analysis of games, as I've already seen some contrast-jacked still shots used to overlay tactical diagrams.

_pferreir_• 12 years ago

I can't see why this was downvoted. I felt exactly the same.

bmm6o2• 12 years ago

Complaining that it refers to the "wrong" football adds nothing to the discussion.

mwfunk2• 12 years ago

As someone else pointed out, there is no one sport that has a monopoly on the term "football". Also, complaining about Americans who refer to American football as "football" often sounds like thinly veiled xenophobia disguised as pedantry, a la "here's one more thing I hate about those stupid Americans..." It's such an odd thing to complain about that it comes across as the tip of a hate iceberg rather than an actual complaint about terminology.

mwfunk2• 12 years ago

Alternately, it's like if someone were to write a color management API for Python, but used the British spelling for "color", and a few commenters on some tech forum got up in arms about the spelling. Equally abrasive (the commenters, that is) for similar reasons (a faint whiff of xenophobia and/or a sense of cultural entitlement).

rmc• 12 years ago

Eh, "Football" is a term that's ambiguous, and different dialects have it refering to different sports.

(Lest you think I'm a yank, I'm Irish, and there's people here who'd say "football" for a different sport (gaelic football)).

dmicah• 12 years ago

There is not one "football" http://en.wikipedia.org/wiki/Football

fmela• 12 years ago

Very neat. It's interesting to see the NFL logo from the corner of the original footage moving around in the stabilized footage. Future feature: automated watermark removal?

aganders3• 12 years ago

As the watermark is static in the original footage, it should be easy to remove before processing.

I'd like to see the players remain, perhaps somewhat transparent, once they are cropped out of the actual shot. Otherwise the zoom/pan just erases them.

euroclydon• 12 years ago

If you sign up for NFL Replay, you can watch "the full 22" videos during the week. It's inexpensive. I hate that they don't show this view on regular broadcasts. It's such a big part of the game, to know what type of coverage the DBs are in when a play starts. To know whether the QB made a good read, you'd have to see the coverage on all of the receivers.

minikites• 12 years ago

I assumed it had something to do with not providing that information to the currently playing teams while they are currently playing, are teams allowed to have "spotters" up in the stands to get an overhead view?

frankchn• 12 years ago

They are - in fact they can take photographs of it and bring it down to the benches: http://www.popularmechanics.com/outdoors/sports/football/0-t...

Apparently they have started replacing the physical photographs with Surface tablets as part of a sponsorship deal with Microsoft as well: http://blog.surface.com/2014/08/surface-ready-nfl-sidelines/

at-fates-hands• 12 years ago

Interesting article from 2011 talking about why the NFL didn't give the "all 22" view to the teams:

http://online.wsj.com/news/articles/SB1000142405297020371620...

"Lonnie Marts, a former linebacker for the Jacksonville Jaguars, says there are thousands of former NFL players who could easily pick apart play-calling and player performance if they had access to this film."

And another one for the conspiracy buffs:

"The NFL makes a handful of plays from the All 22 available on its web site for a fee, but they're often so blurry the players' numbers aren't visible."

clarkm• 12 years ago

I agree. I enjoy football strategy and like seeing different coaching styles in action, but this is hard to pay attention to during a standard broadcast. So I really love the trend toward offering multiple ways to view games. If you haven't seen it already, the "BCS Film Room" edition of last year's National Championship game was one of the best broadcasts I've ever seen.

https://www.youtube.com/watch?v=U-Js6gMlc-U

sharkweek• 12 years ago

I would casually watch games in this format last season just because it's super fascinating -

However, I absolutely fell in love with it rewatching the Super Bowl (Hawks fan) and just following Kam Chancellor - The guy got completely snubbed for MVP

Here's another person's 22 man camera view talking about The Touchdown Canceller's (quite literally perfect) performance:

https://www.youtube.com/watch?v=aE4ng4U3utc

tvon• 12 years ago

Pedantry, but I believe the term is the "All-22".

Edit: I don't know if they offer the All-22 but NBC's Sunday Night Football is always streamed online for free and they offer multiple camera angles. It's pretty nice compared to what other broadcasts offer.

darkstar999• 12 years ago

I'm guessing this is the link

https://gamerewind.nfl.com/nflgr/secure/packages?ttv=0

That's actually pretty cool! Do they do this for college football?

AndyNemmity• 12 years ago

Fantastic Blog, I run an online football game and try to use advanced analytics to generate proper games. Going to rummage through your archive and see if there's anything usable for me, trying to generate accurate games.

Thanks for the work.

Edit: Example, this blog post is amazing for my work. http://phdfootball.blogspot.com/2014/03/classifying-wrs-tes-...

You provide the code you ran(!!!) so I can run it myself against the data I generate, and then validate that I am matching the actual data.

Incredible. Going to continue moving through the article, but that's at least one piece of data I can use to compare in a way I never have before.

Edit 2: http://phdfootball.blogspot.com/2013/10/qbs-dont-have-passin... is interesting, but you have to do any QB completion data with completion % based on distance thrown. QB completion is meaningless without incorporating distance.

Edit 3: http://phdfootball.blogspot.com/2013/09/not-all-fumbles-are-... is data I can incorporate today. Currently I have a simple random recovery generator that is 50/50 due to FO's essentially random remarks regarding fumbles.

Now I will change that to be a formula based on field position. That is direct improvement to my work, thank you!

Edit 4: http://phdfootball.blogspot.com/2013/08/quarterback-rating-i... is again plagued by completion %, which QB rating uses.

Many Qbs improve or decline their QB rating, while actually being better QBs or worse depending on the yardage thrown. I could give many examples of QBs with decent Completion %s that are horrible because they throw short passes.

Not bad, just another note that distance is the foundation of completion % and without that, it's essentially bad data.

One interesting analysis would be plotting the distance thrown on passes over careers, and see if this is a good indicator of improvement.

Edit 5: http://phdfootball.blogspot.com/2013/06/field-position-and-s... is interesting, I distinctly remember that Completion % per yardage drops significantly inside 20 yards. I don't have the data on hand, but that's an interesting wrinkle to consider. I think that may have been FO.

philsnow• 12 years ago

Edit 2: http://phdfootball.blogspot.com/2013/10/qbs-dont-have-passin.... is interesting, but you have to do any QB completion data with completion % based on distance thrown. QB completion is meaningless without incorporating distance.

I had a similar thought; seems that he should bucket at least by remaining yards to get the first down as well as field position. Imagine an 8-yard pass completion after a 1st and ten: many teams will attempt a deep pass, hoping to pick up the first down with a short play in case of an incompletion. That deep pass is a very different play and probably has a different completion likelihood.

Likewise, consider passes from your own 20 vs their 20 (that is, 80 yards of field left before the end zone vs 20): my intuition says that the latter pass has lower likelihood of completion, because the field is shorter and the defenders have less ground to cover.

So, it might actually be that "rhythm" exists, but only in long field positions or only on long drives.

streptomycin• 12 years ago

You forgot to plug your game, http://deeproute.com/ :)

nostromo• 12 years ago

People do image stabilization manually on reddit and the results are amazing.

Ex: http://gfycat.com/CriminalAromaticEsok or http://i.imgur.com/ErQRMrV.gif

http://www.reddit.com/r/ImageStabilization/

wingerlang• 12 years ago

Not fully manual is it?

valgaze• 12 years ago

This technique reminded me of BriefCam- http://briefcam.com/

They figured out how to "summarize" video surveillance footage by attaching a hovering timecode to any object moving through the frame and then later it overlays several hours of footage into one video.

Pretty trippy: https://www.youtube.com/watch?v=fISfDd35sXU

StavrosK• 12 years ago

Oh wow, that's an amazing idea and great execution as well.

csbrooks• 12 years ago

That is really cool.

I wonder how it does when the lighting changes throughout the day?

dougabug• 12 years ago

This seems like just background subtraction. I used this on soccer videos eight years ago using some robust optical flow code from Michael Black to estimate the background motion and suppress it, leaving the foreground (i.e. the player).

kendallpark• 12 years ago

+1 +1 +1

As a programmer, football player, and football coach, I've been searching for something like this for a long time.

I think the biggest advancement in technology would be figuring out what plays are being run and in what scenarios, and analyzing them for patterns and weaknesses. Like knowing that the opposite team runs GAT to the wide side of the field 95% of the time. Or figuring out individual players' tendencies to give more detailed read keys to the defense.

The problem is, I think, that a lot of the critical info is inside the tackle box with the actively engaging linemen. It seems it would be tricker to track ten or so players in such close quarters.

jonifico• 12 years ago

Is this technology on the players' shoulder pads something like Adidas' miCoach?

btown• 12 years ago

Very interesting! With some manual or automatic annotation, this is one step closer to spatiotemporal player data, which opens up whole new boundaries of analysis.

As a point of comparison, there's a private NBA dataset (based on imaging and manual annotation) which shows the locations of shots made, annotated with the player name and whether they failed or succeeded - from this, you can do really cool things like Miller et. al. 2014 [1], who use dimensionality reduction to find patterns in players' preferred shot locations, potentially allowing you to create teams whose offensive abilities are complementary from a spatial perspective.

[1] http://jmlr.org/proceedings/papers/v32/miller14.pdf

Isolating Football Player Movement by Eliminating Camera Motion

Related Stories