Swing Vision AI line calls and Analysis

swupnil said:
@naruto83 haha yea we will definitely try our best to get it working on iPhone 12. One idea I had was that even if it’s not as accurate as iPad Pro (due to processing limitations), we could still release the line calling but allow you to overrule the occasional incorrect calls from your watch? You could even see a replay of the last bounce? I guess my only concern would be how to make sure both players agree to the overrule (there would always be an incentive for one person to disagree unless you are friends and know each other well).

And yes I used to run into the racket smashing issue as well. I just wear my watch on my dominant hand now. Took a couple matches for my non-dominant hand to get used to doing swipe / tap gestures but it’s not too bad! Also you can turn off the Digital Crown scoring in the settings in your watch to avoid phantom gestures :)
Click to expand...


i think if it’s close to 80-90% accuracy it will still be more accurate than most people’s line calls. Hahaha. I would be happy with that. how about get the automatic point scoring up first? If that’s easier? Saves one less manual* thing to do on the court. As the point of all this is to make things automatic* using the AI.

I guess I might need to upgrade my new phone again in the future if I want better accuracy due to the limitations of the processor aye? Hm.... the missus will not be pleased
 
@swupnil Signed up on the site and am so ready and willing to hand over $120, this is an absolutely brilliant idea, wonderful app, and really awesome AI you all are building.

I tried to film my first full match tonight. iPhone 12 Pro on 14.4.2. It was two sets and about 90 minutes, but the app/recording shut off at 16 minutes in. Is this a known bug? Are there any ways to circumvent these bugs to allow the phone to film the entire session w/o any issues? It was on DND! Thank you!
 
@Nostradamus yes we can pretty easily see the other side of the court more clearly than the players can. The latest iPhones and iPads have multiple camera lens that effectively let you "zoom in".

@naruto83 you would be surprised! We still have some work to do before we're more accurate than humans but we're almost there. And the line calling is actually necessary for the auto scoring. There's no way to know who won the point accurately if we're making mistakes on whether the ball was in or out. But as I said, I could see us shipping something where you have the ability to correct things still.

@hpaiste thank you for signing up! That's very strange about the recording ending early. I'm definitely not aware of such a bug and we haven't come across this in our testing. Did you see a low battery warning on the screen when you brought your phone down? That will force the recording to end. These days I put my phone in Airplane mode to completely avoid interruptions of any kind :)

@Demented I'm still not really convinced on why we need to report "TV speeds". If we're getting the average speed throughout the trajectory correct, and you're able to compare your speeds to your opponents / yourself, does it really matter if you have the initial speed vs the average speed? The initial speed is also going to be a lot of effort for us to estimate accurately from a single camera as the drag depends on altitude, temperature, humidity, wind, and so many other variables. I personally don't see the ROI across our entire user base for doing all of this work to get the perfect initial speed if we can get to near-perfect average speed. Correct me if I'm wrong but the benefit to the end user in terms of evaluating and improving their game seems basically the same?
 
Thanks for replying to all of us!!

re: the person believing it’s near the accuracy of a human. 90% accuracy is still lacking versus a person as that 90% is on all calls, not just the hard ones.

re: the person that had it stop in match. I posted a week back about that happening to me, and it had to do with my battery dropping below 20% and iPhone issuing the battery warning, thus interrupting the tracking.

@swupnil Another minor issue that occurs 1-2 times per match - I think predominantly in doubles - is it believes the wrong side of the court is out. It’s odd because it’ll be a single hit (like the serve) in a point and all other “hits” are scored on the correct side. In the editing mode I’ve checked that it’s the correct player hitting the ball, but it still doesn’t flip the court. Maybe I’m doing something wrong.
 
@Demented I'm still not really convinced on why we need to report "TV speeds". If we're getting the average speed throughout the trajectory correct, and you're able to compare your speeds to your opponents / yourself, does it really matter if you have the initial speed vs the average speed? The initial speed is also going to be a lot of effort for us to estimate accurately from a single camera as the drag depends on altitude, temperature, humidity, wind, and so many other variables. I personally don't see the ROI across our entire user base for doing all of this work to get the perfect initial speed if we can get to near-perfect average speed. Correct me if I'm wrong but the benefit to the end user in terms of evaluating and improving their game seems basically the same?
Click to expand...

Maybe I'm wrong and no one else will concur with my assessment but I think you're way off base in terms of the appeal of the product when showing slow numbers. I setup a demo for my club pro a few days ago after the new version hit the street. I showed him a pretty good rally sequence of me putting away an on the run winner down the line. You could tell I slapped the ball pretty hard and then Swingvision reported 43 mph. The shot in reality was like 52 mph. His feedback wasn't about how Swingvision did a great job cutting the film, or the accurate placement on the chart. It was, there's no way that was 42. People who record their tennis matches don't want to see slow numbers, they buy new equipment and train in fact to make those numbers go up. I could go into a further detailed explanation around this but the simple matter is that speed sells....

Second, the other big use for this software is recording junior tennis for highlight reels and film to send to scouts/colleges. We have a big program at the facility I use which includes a PlaySight court for that purpose and I promise you that no one is going to use Swingvision instead if their average shot is being reported 20% below actual. I will continue to use the product but my ability to sell it for your to help adoption isn't going to happen as is.
 
Demented said:
Maybe I'm wrong and no one else will concur with my assessment but I think you're way off base in terms of the appeal of the product when showing slow numbers. I setup a demo for my club pro a few days ago after the new version hit the street. I showed him a pretty good rally sequence of me putting away an on the run winner down the line. You could tell I slapped the ball pretty hard and then Swingvision reported 43 mph. The shot in reality was like 52 mph. His feedback wasn't about how Swingvision did a great job cutting the film, or the accurate placement on the chart. It was, there's no way that was 42. People who record their tennis matches don't want to see slow numbers, they buy new equipment and train in fact to make those numbers go up. I could go into a further detailed explanation around this but the simple matter is that speed sells....

Second, the other big use for this software is recording junior tennis for highlight reels and film to send to scouts/colleges. We have a big program at the facility I use which includes a PlaySight court for that purpose and I promise you that no one is going to use Swingvision instead if their average shot is being reported 20% below actual. I will continue to use the product but my ability to sell it for your to help adoption isn't going to happen as is.
Click to expand...

The speed feature seems to be more of an Ego thing anyway, comparing yourself to tv speeds seems unneeded. I am more concerned with shot tracking accuracy and analyzing my patterns by seeing which is my lowest % shot or when i miss the most etc.
 
@Cailean24 yea there was a bug in one of the versions last week that got the side of the servers wrong but it should be fixed in the latest one. If you still see that, please DM me a link to the session! Maybe you’ve found a challenging case for the A.I. :)

@Demented I wonder in your specific case if that was an issue of tracking error as opposed to our choice of metric. Take a look at the Favorite Points highlight for my match here. I was trying to hit the ball pretty hard and most of the point-ending shots here are in the 50s or 60s. I haven’t edited any of these shots, these are all straight from the A.I. https://swing.tennis/matches/sw1-aQ...ugWDSvF8npVgOhDTtTCn6zWMJkYO-QrDF_GvSm3s6Bh4g

You can tell if it’s a tracking error by editing the shot. If after adjustment of the hit/bounce timing and positions the speeds are very different, this means the A.I. was just off in its predictions.

@Jeremy Lichtman I personally agree with you that speed is, for the most part, a vanity metric. Based on our user interviews across thousands of users, most people agree. We have several D1 programs using the app now and the speeds are good enough for their needs as long as faster shots show a faster speed. Have yet to come across a D1 program that complained about speeds not matching radar guns / TV in the absolute sense.

Of course the A.I. still make mistakes within our choice of speed calculation, and that’s something I’m committed to getting as accurate as possible.
 
The other issue that will fly under the radar so to speak is that the deviance from initial to average grows as the length of the shot increases. So your software is calculating in court balls and short balls more accurately than it's calculating deep baseline hits. A baseline to baseline exchange might average 78 feet in length per shot while short balls or in-court shots will average more like 45 feet.

A 78 foot ball struck at 64 mph will experience a 21% loss of velocity on average resulting in a Swing Vision speed of 51 mph.
A 45 foot ball struck at 64 mph will experience an 11% loss of velocity on average resulting in a Swing Vision speed of 58 mph.

Both balls are struck with the exact same force but differing launch angles and probably topspin amounts but result in very different results. This is purely due to the bounce coming quicker so the measurement is less effected by drag.
 
@Demented that’s a good point and probably the best argument to date for initial speed haha.

I guess my devils advocate argument would be that in that scenario, the 45 foot ball is going to come off the bounce much faster than the 78 foot ball. As such, the higher SwingVision average speed is actually more reflective of how difficult you’re making things for your opponent, which is perhaps more important than how hard you’re striking the ball?
 
No, the ball that bounces short in court will experience a loss of velocity at impact and then continue to lose speed due to drag. It's just not reflected in the numbers since the calculation is until bounce.
 
Last edited:
I guess I didn't realized that the app was calculating average speed vs initial speed you see on TV. Now I feel better because I knew I was ripping some shots and the reported speeds felt slower. I agree that it really doesn't matter and it really is an ego thing. It feels "good" to say I aced you at 120 mph. Makes you feel like you hit a pro level shot but it doesn't really contribute anything to my game other than knowing I can hit the ball hard. I think what is way more helpful are the heat maps and stats like 1st or 2nd server % and shot variation. So if line calling and very detailed match stats can approach 100% I'd be more that happy.

Also @swupnil a feature I'd like to see is PDF export of all the stats/heat map graphics for the match for each player would be super cool.
 
Got a question for you guys. Since the newest update when I export a video a lot of 1st serves are not registering on the overlay. Has anyone else noticed this? Again it's only with exported videos. Not the original source video in the app.
 
I recently got this and installed the app on a Macbook Pro. I'm importing from the Gopro. A couple of things I notice on groundstrokes: 1. The ball tracking is off by up to six feet. Shots landing around the baseline register in the service box. 2. The shot MPH is not accurate (too low most of the time). I rip a forehand and it shows 40 MPH. I hit a slow slice backhand and it shows 50 MPH. A dropshot registered 30 MPH? I have a radar gun and my hard ground strokes are in the 80's. Serve speeds are too slow as well. Nice for stats but would hope there would be better accuracy. I did have my camera setup to ensure that the baseline is was not blocked by the net. 1080p and 30 FPS for the recording settings.
 
720p and 60 fps will yield better results if you can't do 1080 and 60. The frames are more important than the clarity. The height of the camera is also very important. I find that 10 feet is really the minimum in order to get the net low enough that it's below the service line. 12 feet is better. Also every shot is going to be different due to factors such as players or the net occluding the bounce. When the software can clearly see the hit and the bounce, it's fairly consistent. Other systems such as playsight typically rely on 4 or more cameras to ensure that the ball is always visible. With just 1 camera, you're always going to be at the mercy of visibility.
 
Last edited:
Demented said:
720p and 60 fps will yield better results if you can't do 1080 and 60. The frames are more important than the clarity. The height of the camera is also very important. I find that 10 feet is really the minimum in order to get the net low enough that it's below the service line. 12 feet is better. Also every shot is going to be different due to factors such as players or the net occluding the bounce. When the software can clearly see the hit and the bounce, it's fairly consistent. Other systems such as playsight typically rely on 4 or more cameras to ensure that the ball is always visible. With just 1 camera, you're always going to be at the mercy of visibility.
Click to expand...
I will try 60 fps next time. I can live with the speed being off but not the shot placements being off by six feet. Cool tool and love the concept. Very useful if I can tweak things to make it work better and I'm sure it will improve.
 
I was having issues with far side shot placements consistently being off by that much (1-2m), but the update, combined with an actual phone mount (~3m off ground) and the pro version at 1080/60fps really improved it. It still happens, but only a few times per game. Near side only has issues when I’m blocking the view of the bounce.
 
Has anyone ever heard/used this camera mount? I just bought one on Amazon. I love the price and the fact that it's magnetized

f6da1edc2d2ec21b9b1ffd8ad1981cec.png
 
I'd want it to work but I simply can't believe.... especially with the wind nets in the way. One good wind gust and their goes your phone. I made my own(which is what this looks like almost) but I went with an extension pole similar to his and then attached Hercules hooks through the top. I hook the top of the fence or the top chain link and then bungie coord the bottom to create tension.
 
swupnil said:
@Demented that’s a good point and probably the best argument to date for initial speed haha.

I guess my devils advocate argument would be that in that scenario, the 45 foot ball is going to come off the bounce much faster than the 78 foot ball. As such, the higher SwingVision average speed is actually more reflective of how difficult you’re making things for your opponent, which is perhaps more important than how hard you’re striking the ball?
Click to expand...

I'm not sure how PlaySight or Hawkeye report their velocities, but I'm 99% sure that the serve speeds as seen on TV are reported via radar. And radar guns report the fastest velocity observed during the observation interval, which is generally the speed off the racquet. I think that at least for serves, players would want their maximum velocity estimate so they can compare against reported serve speeds on radar-equipped courts.
 
Sorry the delay, everyone!

@jwalk yes we definitely plan to auto generate PDF reports later this year! The serve export issue has been fixed in our 8.7 update that is rolling out on the App Store right now!

@Addxyz yes we are planning to add more stats on winners and points won but right now the shot tracking isn't quite accurate enough to calculate those things automatically. It's close though and likely good enough to at least calculate stats like that in aggregate (even if it's not 100% correct on a per-shot/rally basis). Currently you can see these stats if you also keep score on Apple Watch while recording video on your phone. More about that here.

@jmnk @hpaiste haha yea that speed is super off. You can fix it by tapping the edit shot button and fixing the bounce/hit timing. Our latest 8.7 update also includes some guidelines for optimal accuracy. For example in your video, I would definitely recommend zooming in a lot closer into the court, which will dramatically improve shot placement and speed accuracy.

@Holdfast44ID sorry about that! You'll definitely see a massive improvement if you use 60fps. We actually treat 1080p the same as 720p right now so that matters less. The latest 8.7 update also includes some updated guidelines about how much you zoom in and how high to set up the camera. That should definitely help you optimize the accuracy as well.

@time_fly that makes sense but with a single camera it's a bit challenging to get the initial velocity (we really only have reliable reference points of where the ball was struck from, and where it landed on the court). With 2+ cameras it's pretty trivial since you can accurately determine the position of the ball at every point of the trajectory (hence why PlaySight and Hawkeye can do it). We could potentially get there, for example, if you and your opponent both set up a camera on each end of the court. We also have so much footage now that we could train the A.I. to do this from a single camera but it would come at the cost of decreasing performance of our other metrics like shot placement, etc which I think are more important. Using any sort of physics formula to back solve this is also problematic since the amount of air resistance, wind, etc. all vary depending on the court's altitude and weather conditions.
 
I will say this the air resistance, temperature, and humidity while important only account for a small % of the difference. Distance traveled and drag are by far the biggest components and you could simply add a disclaimer that speeds are back calculated based on average conditions(78 degrees, sea level and 50% humidty). Once you've got a single calculus formula with those inputs it's pretty trivial to adjust the displayed speeds.
 
@Demented thats interesting. Perhaps we could do that and even use the player’s location (if they’ve provided it) to use a more personalized drag coefficient. I’ve definitely seen wind having a big impact if blown in just the right (or wrong, depending on how you look at it) direction but on average it should at least be more reflective of what’s on TV. By the way, you’ve probably already seen this but this is the most thorough academic research I’ve seen on the topic. Pretty cool: https://onlinelibrary.wiley.com/doi/full/10.1002/jst.11#sec1-5-title
 
yeah, I can't help you on wind, that's a biggie that can change constantly.

Tenniswarehouse did their own analysis of drag outside of wind tunnel to get real world conditions. The found lower drag forces at play due to the rotation of the ball that did not occur in the wind tunnel version you referenced.

Tennis Ball Trajectories — The Role of Aerodynamic Drag and Lift in Tennis Shots"

Study to determine how aerodynamic drag and lift forces influence the flight of a tennis ball and to determine the drag and lift coefficients that define these forces.
twu.tennis-warehouse.com

I primarily use https://donthireddy.us/tennis/speed.html to approximate the speeds. It's close enough. Being off +-3 mph isn't as big of a deal compared being low 10-15 on most shots.
His used a research paper that radared the hit at the racket and again at the bounce. Then used calculus to build a formula to approximate the drop off. I've used his calculator(for serves and ground strokes) 100's of times and compared it to not only Playsight but pro footage where they displayed a ground stroke velocity for some reason(usually winner speed). It lines up pretty much every time.
 
I saw a $20 off promo on the swing.vision but no longer see it. How do I get a discounted price? thanks
 
Demented said:
I will say this the air resistance, temperature, and humidity while important only account for a small % of the difference. Distance traveled and drag are by far the biggest components and you could simply add a disclaimer that speeds are back calculated based on average conditions(78 degrees, sea level and 50% humidty).
Click to expand...
you say that air resistance "account for a small % of the difference" but drag "are by far the biggest component". What's the difference between 'air resistance' and 'drag'?

Demented said:
Once you've got a single calculus formula with those inputs it's pretty trivial to adjust the displayed speeds.
Click to expand...
To solve a spinning ball trajectory in the presence of air drag requires solving six first-order differential equations, and they are such that there's no analytical solution. This is nowhere close to 'a single calculus formula'. And there's nothing trivial about solving those.

Demented said:
[...]

I primarily use https://donthireddy.us/tennis/speed.html to approximate the speeds. It's close enough. Being off +-3 mph isn't as big of a deal compared being low 10-15 on most shots.
His used a research paper that radared the hit at the racket and again at the bounce. Then used calculus to build a formula to approximate the drop off. I've used his calculator(for serves and ground strokes) 100's of times and compared it to not only Playsight but pro footage where they displayed a ground stroke velocity for some reason(usually winner speed). It lines up pretty much every time.
Click to expand...
@Demented - why would you say that? Haven't we showed that this calculator gives results that may be off by as much as 15mph compared to a radar gun used at tennis matches? https://tt.tennis-warehouse.com/index.php?threads/frv4-serves.687372/post-15019598

swupnil said:
@Demented thats interesting. Perhaps we could do that and even use the player’s location (if they’ve provided it) to use a more personalized drag coefficient. I’ve definitely seen wind having a big impact if blown in just the right (or wrong, depending on how you look at it) direction but on average it should at least be more reflective of what’s on TV. By the way, you’ve probably already seen this but this is the most thorough academic research I’ve seen on the topic. Pretty cool: https://onlinelibrary.wiley.com/doi/full/10.1002/jst.11#sec1-5-title
Click to expand...

@swupnil - that is actually a really, really old paper. There has been great number of more recent, more detailed work on that topic. I've read quite a few trying to do this https://tt.tennis-warehouse.com/ind...-tennis-serve-trajectory.689458/post-14985755 . If you are interested let me know via PM, I'll send you few links to better research papers.
 
jmnk said:
you say that air resistance "account for a small % of the difference" but drag "are by far the biggest component". What's the difference between 'air resistance' and 'drag'?

To solve a spinning ball trajectory in the presence of air drag requires solving six first-order differential equations, and they are such that there's no analytical solution. This is nowhere close to 'a single calculus formula'. And there's nothing trivial about solving those.

@Demented - why would you say that? Haven't we showed that this calculator gives results that may be off by as much as 15mph compared to a radar gun used at tennis matches? https://tt.tennis-warehouse.com/index.php?threads/frv4-serves.687372/post-15019598
Click to expand...

First off, I obviously meant air density(sea level), stop trolling.

Second, the calculator isn't off, the inputs are off. The 15 mph serve in question was filmed at terrible fps and resolution. I clearly explained that I was guessing on the bounce/hit frames due to the lack of detail. I also called into question whether the radar was correct at that small club. Use real pro footage from tv with radar's behind the server and then apply the calcuators formula and tell me the numbers aren't good enough. You seem to get hung up on 'good enough' as opposed to exact. The cacluator clearly says that it's +-3% even when the inputs are perfect. It's obviously going to be a much wider range when you're basically guessing.
 
Thanks, @jmnk . That's very helpful. I realized how old the paper was after checking it again haha. My bad. Yikes.

In any case, the nuances of this all gets me back to wondering whether it's worth the effort to report initial speeds if we can (some day soon) get nearly perfect average speeds, which you can compare against yourself/opponents/anyone in the world...

Maybe we have to get a deal to start using SwingVision at all ATP matches and report average speeds on TV. Then everyone will be happy :)
 
swupnil said:
Thanks, @jmnk . That's very helpful. I realized how old the paper was after checking it again haha. My bad. Yikes.

In any case, the nuances of this all gets me back to wondering whether it's worth the effort to report initial speeds if we can (some day soon) get nearly perfect average speeds, which you can compare against yourself/opponents/anyone in the world...

Maybe we have to get a deal to start using SwingVision at all ATP matches and report average speeds on TV. Then everyone will be happy :)
Click to expand...
I don't suppose your software can accurately tell in what frame the ball crossed the net? Hit to net is a much shorter distance which leads to an average velocity measurement that is much closer to initial.
 
@Demented I was actually thinking about that as the most viable short term solution if we really want to solve this. Without disclosing too much, we have a way to accurately determine the exact frame and x/y/z when the ball crossed the net, if we want to. But it's a lot of work for us to train the A.I. to do this so I'm not really sure it's worth the effort.

(I think) this would actually allow us to pretty much perfectly estimate the net aerodynamic force as we would have 3 reference points throughout the trajectory so that would be enough the constrain the trajectory.

An additional argument for pursuing this would be that we could report net clearance, which is actually a very important metric that pretty much every high performance coach has asked me to add.
 
Demented said:
First off, I obviously meant air density(sea level), stop trolling.

Second, the calculator isn't off, the inputs are off. The 15 mph serve in question was filmed at terrible fps and resolution. I clearly explained that I was guessing on the bounce/hit frames due to the lack of detail. I also called into question whether the radar was correct at that small club. Use real pro footage from tv with radar's behind the server and then apply the calcuators formula and tell me the numbers aren't good enough. You seem to get hung up on 'good enough' as opposed to exact. The cacluator clearly says that it's +-3% even when the inputs are perfect. It's obviously going to be a much wider range when you're basically guessing.
Click to expand...
Don't be disrespectful.
You clearly want to believe that the 'frame calculated results' are accurate, as much as radar guns measured speeds are. I should have indeed stopped discussing this with you weeks ago as there's nothing that can deter you from your beliefs.
 
I just plugged some values into the calculator. Due to drag being pretty front loaded, the deviation is still on average around 9-12% depending on speed(when assuming most shots are measured at 40 feet or less). That's clearly better than 9 to 25% though.
 
Not to poke the bear so to speak, but you're already adept at using heuristic processing to figure out stuff. Couldn't you just feed the AI a couple thousand examples of initial velocity and distance traveled and time(no video needed) and have it build a heuristic algorithm to tell you the speed on demand? based on the 2 data points you already have?
 
swupnil said:
Sorry the delay, everyone!

@jwalk yes we definitely plan to auto generate PDF reports later this year! The serve export issue has been fixed in our 8.7 update that is rolling out on the App Store right now!

@Addxyz yes we are planning to add more stats on winners and points won but right now the shot tracking isn't quite accurate enough to calculate those things automatically. It's close though and likely good enough to at least calculate stats like that in aggregate (even if it's not 100% correct on a per-shot/rally basis). Currently you can see these stats if you also keep score on Apple Watch while recording video on your phone. More about that here.

@jmnk @hpaiste haha yea that speed is super off. You can fix it by tapping the edit shot button and fixing the bounce/hit timing. Our latest 8.7 update also includes some guidelines for optimal accuracy. For example in your video, I would definitely recommend zooming in a lot closer into the court, which will dramatically improve shot placement and speed accuracy.

@Holdfast44ID sorry about that! You'll definitely see a massive improvement if you use 60fps. We actually treat 1080p the same as 720p right now so that matters less. The latest 8.7 update also includes some updated guidelines about how much you zoom in and how high to set up the camera. That should definitely help you optimize the accuracy as well.

@time_fly that makes sense but with a single camera it's a bit challenging to get the initial velocity (we really only have reliable reference points of where the ball was struck from, and where it landed on the court). With 2+ cameras it's pretty trivial since you can accurately determine the position of the ball at every point of the trajectory (hence why PlaySight and Hawkeye can do it). We could potentially get there, for example, if you and your opponent both set up a camera on each end of the court. We also have so much footage now that we could train the A.I. to do this from a single camera but it would come at the cost of decreasing performance of our other metrics like shot placement, etc which I think are more important. Using any sort of physics formula to back solve this is also problematic since the amount of air resistance, wind, etc. all vary depending on the court's altitude and weather conditions.
Click to expand...

In the future, can you please quote the posts as well so I can easily follow what you're replying to? Thanks
 
@Demented yes we can do that but calculating the initial velocity for each example is what would take us a long time. Again can't elaborate on how we can get the true initial velocity, but we can. It is very labor intensive though. That's the bottle neck.

To be fair, this is equally as labor intensive as getting the net position/timing. We'll likely try both down the road and go with whichever produces more accurate results in terms of determining the full trajectory. My guess is actually that the net idea will be more reliable since intuitively that is easier for even a human to estimate visually from a single camera (relative size of the ball/net/lines + movement of the ball's shadow help a lot visually, and a neural net will take advantage of those cues).
 
@jwalk yes editing the shot actually automatically incorporates that segment into our A.I.! There is some manual verification/QA to filter out trolling but yes, it’s massively helpful!
 
swupnil said:
@jwalk yes editing the shot actually automatically incorporates that segment into our A.I.! There is some manual verification/QA to filter out trolling but yes, it’s massively helpful!
Click to expand...
You should create a score board for who corrects the most shots.. I'd be interested to see where I rank....
 
I'm very excited to use SwingVision... just waiting for my iPhone 12 to get here to fully benefit from it.

I'm not sure if this feature is already available in the app, but I think it would be very cool if there could be a way the app could tie and pool together every metric it analyzes (i.e. every ground stroke motion, serve motion, spin, speed, ball placement, footwork, unforced errors, shot selection, wins/losses over other players who use the app if you play a match with them, etc) and gives you an estimated NTRP/ITN/UTR ranking based on all that data.

@swupnil Would this be possible? Again if this is already in the app then I'll feel dumb haha
 
Hey so what do you recommend for mounting at courts with nets? The clay courts I play at don't have the normal metal fences, just posts with nets running from each post.
 
@initialize this isn’t in the app yet but once it gets more accurate and more people use it, this is the long term vision, yes. Theoretically we will have access to the most granular data possible so match making will become super personalized and lead to really fun, competitive matches :)

@antony that’s interesting. You should be able to still hang a QM-1 mount at the top of the net, unless it’s not that sturdy, in which case a tripod may be the safest option.
 
swupnil said:
@initialize this isn’t in the app yet but once it gets more accurate and more people use it, this is the long term vision, yes. Theoretically we will have access to the most granular data possible so match making will become super personalized and lead to really fun, competitive matches :)

@antony that’s interesting. You should be able to still hang a QM-1 mount at the top of the net, unless it’s not that sturdy, in which case a tripod may be the safest option.
Click to expand...
Awesome, thank you man.

another idea as well... this might be really dumb/hard to implement, so go easy on me here lol. It would be cool if you guys analyzed all the top pro players (and past/retired pros) individual strokes, and then maybe the app could have some sort of analyzer function that compares each of your strokes to a pro player with a similar stroke motion. For instance, the app sees that your forehand is most similar to Federer, and then there’s a side-by-side slow motion video comparing your forehand to his that shows the similarities/differences of the stroke, and ways you can improve your forehand more (I.e the app notices you’re too late or early hitting the ball, your wrist is too tight, the racket is too close to your body, etc)...

Maybe I’m just dreaming here, but this would be SICK if you guys could do this
 
@initialize its gonna happen! Will take some time to get there but yes we will do it. Also even just comparing to yourself, say your forehand on break point vs your “ideal” forehand in your lessons/practices. Same for serves. It’s gonna be epic :)
 
swupnil said:
@initialize its gonna happen! Will take some time to get there but yes we will do it. Also even just comparing to yourself, say your forehand on break point vs your “ideal” forehand in your lessons/practices. Same for serves. It’s gonna be epic :)
Click to expand...


Heck yes!!! Dude, thank you for making the best app ever
 
So I have been dutifully uploading my sessions for a few weeks now and so far I am a little disappointed in the accuracy and shot recognition (especially on the serve). My kick serves tend to read as flat serves even if the spin is obvious to my human eye. I did get an iphone now so I will see if using the app to film and 60fps will make a big improvement over importing footage.

On the positive side, I have got some use of the statistics to see I serve much better on the deuce side than the ad side so I can work on that specifically.
 
@Jeremy Lichtman thanks for the feedback! The 60fps definitely makes a big difference and the quality of the latest iPhones help a lot too. We’ve also added some camera height & zoom guidelines in the latest 8.7 version that help you set up your recording for optimal accuracy.

If you’re still seeing issues, please use the Edit Shot feature when playing your video (button in the top left) to fix up the spins of your serves. This will help the A.I. do a better job in future updates of the app :)
 
First this app is truly amazing. Sure there are some issues but just editing down a 3 hour match to 45 minutes alone is great. My main problem with the app is I thought I would look like a 4.0 level Federer. As it turns out it is gross flattery to say I look like a 2.5 Jack Black.

swupnil said:
@time_fly that makes sense but with a single camera it's a bit challenging to get the initial velocity (we really only have reliable reference points of where the ball was struck from, and where it landed on the court). With 2+ cameras it's pretty trivial since you can accurately determine the position of the ball at every point of the trajectory (hence why PlaySight and Hawkeye can do it). We could potentially get there, for example, if you and your opponent both set up a camera on each end of the court. We also have so much footage now that we could train the A.I. to do this from a single camera but it would come at the cost of decreasing performance of our other metrics like shot placement, etc which I think are more important. Using any sort of physics formula to back solve this is also problematic since the amount of air resistance, wind, etc. all vary depending on the court's altitude and weather conditions.
Click to expand...

People have focused on the "impact speed" versus the "average speed" between impact and the first bounce. But I would suspect with many lobs or heavy top spin shots the issue is mostly that you are measuring a straight line of the same height where the ball is hit and where it lands and dividing that number by the time. In other words I would think the lack of measuring the "height distance" the ball moves through accounts for at least as much of the difference. Does that make sense? Or does the software try to track how high the ball goes?

In any event I think the speed calculations are fine. I am mainly concerned with understanding which hits are relatively faster than others as opposed to trying to get accurate absolute speeds. I would like to know if generally lobs or heavy top spin shots will register as slower than flat shots of the same absolute speed.
 
You must log in or register to reply here.

Similar threads

D
problems with Swing Vision video playback
Replies
3
Views
1K
waterpro
W
A
Swing Vision: great app, appalling customer support
Replies
3
Views
904
DrTompu
DrTompu
ruhil
Using Maps and Data Visualization to Understand Tennis
Replies
0
Views
635
ruhil
ruhil
N
how to best measure serve speed, and compare to pros...
Replies
28
Views
9K
Injured Again
I
S
Electronic Line Calls
Replies
2
Views
932
West Coast Ace
West Coast Ace
Back
Top