A case for building football analysis tools with the people who actually read the pitch
I watched the Roboflow sports demo. Bounding boxes track players across the frame. Jersey numbers get read. Team colours get classified. A radar view maps coordinates onto a 2D pitch. It is, by any technical standard, impressive.
It also tells me almost nothing about what is actually happening in the match.
This is not a criticism. It is an observation about where the product sits on the value chain, and more importantly, where it could go if the right people are in the room when decisions get made about what to build next.
The gap between tracking and understanding
Computer vision can now tell you where every player is at any given moment. What it cannot yet tell you, at least not in the demos I have seen, is why they are there. And in football, the why is everything.
Consider the concept of half-spaces. These are the vertical channels between the centre of the pitch and the wide areas, roughly aligned with the inside edges of the penalty box extended to the halfway line. They do not appear on any pitch markings. No camera operator frames them. No bounding box will ever detect one. And yet they are the most contested territory in modern football. The best teams in the world, the ones coached by people who have spent decades thinking about positional play, build their entire attacking structure around occupying and exploiting these zones.
When Pep Guardiola's Barcelona invented the modern version of this, the insight was simple: a player receiving the ball in the half-space can see the goal, the near post runner, the far post runner, and the wide option simultaneously. A player receiving the ball in a wide area can only really go forward or backwards. The half-space is where decisions multiply. It is where football gets interesting.
Now. Can a YOLO model detect that a midfielder has drifted into the left half-space? Technically, yes, if you define the zone coordinates and check whether the player's position falls within them. But that is not analysis. That is a lookup. The analysis is understanding what that movement created: did it drag the opposing centre-back out of position? Did it open a passing lane to the striker? Did the fullback behind him overlap into the space he vacated, creating a 2v1 on the wing? Did the opposite winger tuck inside to form a triangle that the defence now cannot cover without leaving someone free?
These are the questions that coaches, analysts, and anyone who truly watches football are trying to answer. The technology should be built to help them answer faster, not to answer different, simpler questions that happen to be easier to compute.
What Saturday's final will actually be about
Arsenal and PSG meet in Budapest on Saturday for the Champions League final. PSG are defending champions. Arsenal are in their first final in twenty years. On paper, two elite squads. In practice, two completely different philosophies of how football should be played.
Arsenal under Arteta are a team built on compactness and collective discipline. They conceded six goals in the entire Champions League campaign. Six. Their defensive block is not reactive; it is architecturally designed. The distances between centre-backs, the positioning of the double pivot, the triggers that tell the wide players when to tuck inside versus when to press high. All of it is choreographed. When Arsenal defend well, the spaces between their lines are so compressed that opponents run out of ideas, not because they lack talent, but because they cannot find a passing angle that leads anywhere dangerous.
PSG are the inverse. They want to stretch the pitch, rotate positions, and create numerical advantages through movement rather than structure. Kvaratskhelia drifts inside from the left. Dembélé isolates defenders on the right. The full-backs push high to provide width. The whole system is designed to pull defensive shapes apart through constant positional interchange, so that when the ball arrives in a dangerous area, someone is free because the defenders could not track all the rotations.
Here is what a football analyst would want to see from a tool built on Roboflow's detection layer:
Defensive shape compression over time. Not just a heatmap, but a dynamic measurement of the distance between Arsenal's lines (defence to midfield, midfield to attack) at different phases of play. When PSG have possession in the middle third, how tight does Arsenal's block get? When it gets tightest, where does PSG find the breakthrough, if they find one at all?
Triangle detection. Football is played in triangles. Every good passing sequence involves at least one. When a player receives the ball, the two nearest teammates should form a triangle that gives the ball carrier at least two forward or lateral options. A tool that can identify and visualise triangles in real time, and flag when they break down, would be genuinely useful to coaches. It would show where a team's structure holds and where it frays.
Half-space occupation frequency and outcome. How often does each team get a player on the ball in the half-space? What happens next? Does it lead to a shot, a key pass, a dribble into the box, or a safe pass backwards? This is the kind of contextual data that currently requires a human analyst to code manually from video. Automating even part of it would save professional clubs hundreds of hours per season.
Overload detection. When PSG rotate three players to one side of the pitch to create a 3v2, can the system recognise that numerical advantage in real time and flag it? Can it then track whether the overload led to a crossing opportunity, a shot, or a turnover? The pattern itself is simple geometry. The value is in connecting the geometry to the outcome.
Press triggers and press resistance. When Arsenal press high, what triggers it? The opponent's first touch? A backwards pass? A specific player receiving the ball? And when PSG play through the press, which player combinations are they using? This is pattern recognition at a level that current bounding-box tracking does not attempt, but could, with the right guidance on what patterns to look for.
The product problem is not technical. It is conversational.
Roboflow has solved the hard part. The detection models work. The tracking works. The pitch mapping works. What has not been built yet is the interpretive layer that makes any of it useful to the people who would actually pay for it: coaches, performance analysts, scouting departments, and media companies producing tactical content.
And that layer cannot be built by engineers alone. Not because engineers lack intelligence, but because they lack the thousands of hours of pattern recognition that come from watching, playing, coaching, and obsessing over football. An engineer sees a cluster of bounding boxes. A football analyst sees a pressing trap. An engineer sees a player moving laterally. A football analyst sees a centre-back stepping into midfield to play a line-breaking pass, which changes the entire structure of the attack.
My recommendation, for whatever it is worth, is to go and find those people. Sit with coaches. Sit with analysts. Sit with the obsessives who pause matches every thirty seconds to study where the left-back is relative to the number ten. Ask them what they wish they could see but currently cannot. Build from their vocabulary, not from the model's output labels.
The technology is ready. The question is whether anyone is asking it the right questions.
I'm a technical recruiter turned career advisor who spends their days helping engineering students articulate what they have actually done, and their evenings arguing about whether the false nine is a position or a philosophy. Both activities, it turns out, are fundamentally about the same thing: seeing what is really there, not what the framework tells you should be.
A case for building football analysis tools with the people who actually read the pitch
I watched the Roboflow sports demo. Bounding boxes track players across the frame. Jersey numbers get read. Team colours get classified. A radar view maps coordinates onto a 2D pitch. It is, by any technical standard, impressive.
It also tells me almost nothing about what is actually happening in the match.
This is not a criticism. It is an observation about where the product sits on the value chain, and more importantly, where it could go if the right people are in the room when decisions get made about what to build next.
The gap between tracking and understanding
Computer vision can now tell you where every player is at any given moment. What it cannot yet tell you, at least not in the demos I have seen, is why they are there. And in football, the why is everything.
Consider the concept of half-spaces. These are the vertical channels between the centre of the pitch and the wide areas, roughly aligned with the inside edges of the penalty box extended to the halfway line. They do not appear on any pitch markings. No camera operator frames them. No bounding box will ever detect one. And yet they are the most contested territory in modern football. The best teams in the world, the ones coached by people who have spent decades thinking about positional play, build their entire attacking structure around occupying and exploiting these zones.
When Pep Guardiola's Barcelona invented the modern version of this, the insight was simple: a player receiving the ball in the half-space can see the goal, the near post runner, the far post runner, and the wide option simultaneously. A player receiving the ball in a wide area can only really go forward or backwards. The half-space is where decisions multiply. It is where football gets interesting.
Now. Can a YOLO model detect that a midfielder has drifted into the left half-space? Technically, yes, if you define the zone coordinates and check whether the player's position falls within them. But that is not analysis. That is a lookup. The analysis is understanding what that movement created: did it drag the opposing centre-back out of position? Did it open a passing lane to the striker? Did the fullback behind him overlap into the space he vacated, creating a 2v1 on the wing? Did the opposite winger tuck inside to form a triangle that the defence now cannot cover without leaving someone free?
These are the questions that coaches, analysts, and anyone who truly watches football are trying to answer. The technology should be built to help them answer faster, not to answer different, simpler questions that happen to be easier to compute.
What Saturday's final will actually be about
Arsenal and PSG meet in Budapest on Saturday for the Champions League final. PSG are defending champions. Arsenal are in their first final in twenty years. On paper, two elite squads. In practice, two completely different philosophies of how football should be played.
Arsenal under Arteta are a team built on compactness and collective discipline. They conceded six goals in the entire Champions League campaign. Six. Their defensive block is not reactive; it is architecturally designed. The distances between centre-backs, the positioning of the double pivot, the triggers that tell the wide players when to tuck inside versus when to press high. All of it is choreographed. When Arsenal defend well, the spaces between their lines are so compressed that opponents run out of ideas, not because they lack talent, but because they cannot find a passing angle that leads anywhere dangerous.
PSG are the inverse. They want to stretch the pitch, rotate positions, and create numerical advantages through movement rather than structure. Kvaratskhelia drifts inside from the left. Dembélé isolates defenders on the right. The full-backs push high to provide width. The whole system is designed to pull defensive shapes apart through constant positional interchange, so that when the ball arrives in a dangerous area, someone is free because the defenders could not track all the rotations.
Here is what a football analyst would want to see from a tool built on Roboflow's detection layer:
Defensive shape compression over time. Not just a heatmap, but a dynamic measurement of the distance between Arsenal's lines (defence to midfield, midfield to attack) at different phases of play. When PSG have possession in the middle third, how tight does Arsenal's block get? When it gets tightest, where does PSG find the breakthrough, if they find one at all?
Triangle detection. Football is played in triangles. Every good passing sequence involves at least one. When a player receives the ball, the two nearest teammates should form a triangle that gives the ball carrier at least two forward or lateral options. A tool that can identify and visualise triangles in real time, and flag when they break down, would be genuinely useful to coaches. It would show where a team's structure holds and where it frays.
Half-space occupation frequency and outcome. How often does each team get a player on the ball in the half-space? What happens next? Does it lead to a shot, a key pass, a dribble into the box, or a safe pass backwards? This is the kind of contextual data that currently requires a human analyst to code manually from video. Automating even part of it would save professional clubs hundreds of hours per season.
Overload detection. When PSG rotate three players to one side of the pitch to create a 3v2, can the system recognise that numerical advantage in real time and flag it? Can it then track whether the overload led to a crossing opportunity, a shot, or a turnover? The pattern itself is simple geometry. The value is in connecting the geometry to the outcome.
Press triggers and press resistance. When Arsenal press high, what triggers it? The opponent's first touch? A backwards pass? A specific player receiving the ball? And when PSG play through the press, which player combinations are they using? This is pattern recognition at a level that current bounding-box tracking does not attempt, but could, with the right guidance on what patterns to look for.
The product problem is not technical. It is conversational.
Roboflow has solved the hard part. The detection models work. The tracking works. The pitch mapping works. What has not been built yet is the interpretive layer that makes any of it useful to the people who would actually pay for it: coaches, performance analysts, scouting departments, and media companies producing tactical content.
And that layer cannot be built by engineers alone. Not because engineers lack intelligence, but because they lack the thousands of hours of pattern recognition that come from watching, playing, coaching, and obsessing over football. An engineer sees a cluster of bounding boxes. A football analyst sees a pressing trap. An engineer sees a player moving laterally. A football analyst sees a centre-back stepping into midfield to play a line-breaking pass, which changes the entire structure of the attack.
My recommendation, for whatever it is worth, is to go and find those people. Sit with coaches. Sit with analysts. Sit with the obsessives who pause matches every thirty seconds to study where the left-back is relative to the number ten. Ask them what they wish they could see but currently cannot. Build from their vocabulary, not from the model's output labels.
The technology is ready. The question is whether anyone is asking it the right questions.
I'm a technical recruiter turned career advisor who spends their days helping engineering students articulate what they have actually done, and their evenings arguing about whether the false nine is a position or a philosophy. Both activities, it turns out, are fundamentally about the same thing: seeing what is really there, not what the framework tells you should be.