Sports analysis presents a configuration few domains combine: visually identical objects, a known geometric reference and a requirement bearing on events rather than on positions. It is the hardest case of identity tracking and the best equipped.
This article sets out what this context demands. It extends the article on video annotation for autonomous driving.
What distinguishes this context
Five properties characterise sports video annotation. Objects on the same team are visually identical, kit, colour and build not separating them. The pitch is a known reference, its dimensions and markings supplying an exploitable geometry. The rules bound the possible configurations, a player not being able to be anywhere or do anything. The number of actors is fixed and known in advance, a strong constraint existing in no other vertical. And the expected deliverable often concerns events and statistics rather than raw trajectories. One important practical consequence follows. The fourth property supplies a check no other domain offers, the number of active tracks having to match the number of players on the pitch, an immediate control without equivalent in video annotation.The similarity problem
This property makes it the hardest case of identity tracking. Two players on the same team are distinguished only by shirt number and by context. Four difficulties follow. The number is legible only from certain angles, a player facing forward or in profile not showing it. Groupings produce numerous brief mutual occlusions. Physical contact makes separating the silhouettes ambiguous. And changes of direction are abrupt, which invalidates pairing by trajectory continuity. One important observation follows. Those four difficulties compound in the passages of play that most interest analysis, corners and scrums notably, which means difficulty grows exactly where value concentrates.The two levels of identity
This distinction clarifies the scoping of a sports video annotation project. It is rarely stated. Track identity guarantees that an object keeps an internal number across the sequence. Named identity attaches that track to a designated player, information external to the image. Three consequences follow for a video annotation project. The first level suffices for measuring distances covered and space occupation. The second is indispensable to any individual statistic and it costs considerably more. And the second is sometimes derived from the first, a correctly tracked path needing to be named only once. One practical consequence follows. That derivation is this vertical’s principal saving lever, effort going to track continuity rather than to repeated recognition of players.What the pitch supplies
This free geometric information distinguishes the vertical. The markings permit computing the transformation between the image and the pitch plane. Four exploitations follow. Conversion to pitch coordinates, which makes positions comparable between cameras and between sequences. Plausibility checking, a position outside the boundaries signalling an error. Trajectory regularisation, a movement being expressed in metres per second rather than in pixels. And speed checking, a physically impossible value signalling an identity switch. One important observation follows. The last point supplies a switch detector more reliable than anything the other verticals permit, since an athlete’s maximum speed is known and tightly bounded.What the rules supply
A second source of constraint is available. It stays underused in video annotation. The rulebook bounds the configurations play can produce. Three exploitations follow. Counting, a number of players above the permitted squad signalling a duplicated track. Forbidden zones, a goalkeeper outside their area in certain sports contradicting an identity attribution. And stoppages, moments where positions stabilise and where a reliable reassignment becomes possible. One practical consequence follows for a video annotation project. The third exploitation is the most useful in practice, stoppages offering resynchronisation points where every identity can be verified and corrected before play resumes.The ball as an object apart
One object concentrates a difficulty out of proportion to its size and it deserves separate treatment. The ball is small, fast and frequently hidden, and its position nonetheless determines most of the analysis. Four difficulties characterise it. Its speed produces motion blur that spreads its outline over several pixels. Its apparent size varies strongly with distance to the camera. It disappears regularly, masked by players or out of frame. And its trajectory is ballistic rather than continuous, which invalidates linear interpolation on aerial phases. One important practical consequence follows for a video annotation project. Those four difficulties justify handling the ball as a distinct line with its own rate, often higher than the players’, rather than including it in the general tracking at the same regime.The expected deliverable
An essential difference separates this vertical from the others. The client rarely expects trajectories and often expects events. Four families of deliverable occur. Movement statistics, distances, speeds and zones occupied, derived directly from the trajectories. Game events, passes, shots and duels, which presuppose a distinct temporal annotation. Tactical configurations, lines and blocks, which concern sets of players rather than individuals. And notable sequences, clips intended for editing or analysis. One observation follows. The second family requires a sporting competence that annotating positions does not, which makes it a distinct line rather than an extension of the tracking work.Annotating events
Four decisions structure this video annotation task. They are taken with someone who knows the sport. The event vocabulary, whose granularity determines everything else. The instant convention, an event being bounded by a start and an end or reduced to a point. Attribution, an event being attached to one player, to two or to none. And the handling of ambiguous cases, an interrupted or contested action calling for a rule. One important practical consequence follows. The second decision produces the commonest disagreement between annotators, the start of a pass being placeable at the gesture, at the contact or at the ball’s departure, which requires a written and illustrated definition rather than an assumed shared understanding.Building the event reference
One operation conditions the consistency of all event annotation and it is conducted with someone who knows the sport. Five elements compose an event’s entry. The definition, expressed in observable criteria rather than in the player’s supposed intention. The reference instant, start, contact or end, settled explicitly. The associated confusions, neighbouring events to distinguish it from and the criterion separating them. Illustrated examples, video clips showing the typical case and the edge cases. And what to do in case of doubt, a dedicated class rather than a default choice. One practical consequence follows for a video annotation project. This reference takes a few hours of someone who knows the sport and it determines the consistency of thousands of annotations, which makes it the best-return investment in this type of project.Camera configurations
Three arrangements occur in sports video annotation. They do not call for the same work. The fixed camera covering the whole pitch, which gives a complete view and low resolution on each player. The broadcast camera, which follows the action and permanently loses sight of some players. And the multi-camera system, which covers the whole pitch at sufficient resolution at the price of registration between views. One important observation follows for a video annotation project. The second configuration is the most available and the least suitable, a player leaving frame and returning posing a re-identification problem the first does not have.Pose as a complementary deliverable
One task frequently joins tracking and it changes the nature of the work. Annotating joint points permits analysing a gesture rather than a displacement. Three uses justify it. Technical analysis, where the quality of a movement matters more than the position. Injury prevention, where repeated postures signal a risk. And fine action recognition, a shot and a pass being distinguished by the gesture rather than by the player’s trajectory. Three difficulties accompany it. The load per object is far higher, a pose comprising a dozen points where a box requires two. Occlusions make some points unobservable, which requires a convention of estimation or omission. And the resolution available on a distant player does not permit placing the points precisely. One practical consequence follows for a video annotation project. This task combines well only with a camera configuration offering sufficient resolution, which reserves it for close arrangements rather than wide views.What quality control must verify
Five video annotation checks exploit the constraints specific to this vertical. Counting by team, whose departure from the expected squad immediately signals a defect. Speed in pitch coordinates, whose exceeding of the physically possible value reveals a switch. Continuity at stoppages, moments where every identity must be verifiable. Consistency between position annotation and event annotation, a shot attributed to a player absent from the zone constituting a contradiction. And track duration, excessive fragmentation signalling a re-identification failure after occlusion. One practical consequence follows. Those five checks are automatable and they exploit constraints the sport supplies for free, which gives this vertical a verification arrangement more powerful than that of most video annotation projects.The other sports and their constraints
Four families are distinguishable and they do not present the same difficulties. Team field sports, football and rugby notably, where the extended surface and the number of actors dominate. Indoor sports, basketball and handball, where density is higher and the surface smaller, which improves resolution per player. Individual sports, athletics and swimming, where identity poses no difficulty and where the issue is measuring gestures. And racket sports, where the number of actors is minimal and where the ball is the principal object. One observation follows. The last two families escape almost entirely the difficulties set out in this chapter, which makes them markedly more accessible entry points for a video annotation provider approaching this vertical.What this vertical costs
Four factors determine the video annotation load. The number of actors, fixed and known, which makes sizing more predictable than elsewhere. The density of groupings, which concentrates the difficulty into a fraction of the duration. The requirement for named identity, which multiplies the cost against track identity. And event annotation, a distinct line requiring another competence. One observation follows. The first factor is a rare advantage, a match having a duration and a squad known in advance, which permits fixed-price costing far more reliably than in the other verticals.The project calendar
A rhythm constraint distinguishes this vertical and it affects organisation as much as method. The value of an analysis falls rapidly after the fixture. Three consequences follow. Delivery time often takes priority over refinement, an analysis available before the next training session being worth more than a fuller one a week later. The load concentrates on match days, which produces activity peaks incompatible with smoothing. And the season structures the year, an intense period alternating with a quiet one usable for building reference corpora. One important observation follows for a video annotation provider. Those three constraints favour an organisation able to absorb peaks and to work to short windows, a capability that distinguishes more than superior annotation performance does.What sensor data changes
A complementary source is frequently available and it modifies annotation’s role. Locating devices worn by players supply positions without passing through the image. Three consequences follow. Identity is resolved by construction, each device being attached to a named player. Positions are supplied in pitch coordinates, which removes the need for a transformation. And coverage is complete, no player leaving the range of a worn device. Three limits persist. The ball is generally not equipped, which leaves the central object to the image. Pose and gesture are not captured by a position alone. And devices are not permitted in every competition nor worn by opponents. One important practical consequence follows for a video annotation project. That last limit is decisive, an analysis of the opposing team being able to rest only on the image, which sustains annotation demand even where sensors are deployed.What this context does not solve
Four limits deserve stating. Two players in prolonged contact stay inseparable, the data not carrying the information. Shirt numbers are not permanently legible, which leaves intervals where named identity rests on continuity alone. Sports with no marked pitch do not benefit from the geometric reference. And intense contact sports present an occlusion density no method entirely handles. That second limit deserves emphasis. It means named identity depends entirely on tracking quality, an interrupted track having to be renamed manually, which makes continuity the central issue rather than recognition.The first match to handle
Four criteria compose a representative pilot batch in sports analysis. One passage of open play, where players are dispersed and tracking is easiest. One set piece with a grouping, corner or scrum, which concentrates occlusions and tests the convention. One stoppage, which verifies that identity resynchronisation is practicable. And one sequence containing the most frequent events, which tests the reference before generalising it. One observation follows. That batch represents a few minutes of a match, it supplies the throughput per phase type and it reveals the disagreements over event definitions, which makes it the only serious basis for a costing.Approaching a sports analysis project
Five questions scope this type of video annotation project. Track identity or named identity. That answer changes the cost by a substantial ratio. Does the deliverable concern positions or events. That answer determines the competence required. Which camera configuration. That answer conditions the feasibility of re-identification. Is the pitch reference exploitable. That answer determines access to the automatic checks. And is sporting competence available. Its absence caps the quality of any event annotation. Those five answers determine the scope, the method and the control. Asking them before starting avoids a corpus of positions where the client expected statistics.The question that frames the project
One question precedes all the others and it determines the cost by a substantial ratio. Does the expected statistic concern the team or the player. A collective answer, space occupation or density by zone, is satisfied by track identity and sometimes by a simple count. An individual answer, distance covered per player or number of passes, requires named identity and everything it presupposes. That question is asked in one sentence, it is rarely asked, and its answer determines whether the project is ordinary tracking or the hardest case in video annotation.What this chapter teaches
One cross-cutting observation deserves closing this examination. This vertical combines maximum tracking difficulty with the most powerful control arrangement. Three findings compose it. The similarity of objects makes pairing by appearance inoperative, which throws everything onto continuity and context. The constraints of the pitch and the rules supply automatic checks no other domain offers for free. And those two properties offset one another, annotation difficulty being high and error detectability being just as high. That finding matches the one the series on row crops established in an unrelated domain: where structural information independent of appearance exists, it is worth more than a refinement of the visual model, and it serves control as much as production.Who buys this and what they already have
One market observation deserves stating, since this vertical is more crowded than the others in this cluster. Established providers already supply full-service analysis to top-tier clubs and leagues. Three segments stay open to a video annotation provider. Lower-division and youth clubs, whose budgets exclude the full-service offering and whose needs are narrower. Technology companies building analysis products, which need annotated corpora rather than analysis and which are the natural clients for an annotation provider. And sports outside the main commercial ones, where no established offering exists and where the reference work remains to be done. One practical consequence follows. The second segment is the most natural fit, since it buys exactly what an annotation provider produces, and it is reached by approaching the companies building the tools rather than the organisations using them. One observation completes that. A club buys analysis; a product company buys corpora, which is a different sale and a considerably more durable relationship.Common mistakes
These failures recur often enough that naming them is usually enough to avoid them.- Requiring named identity where track identity would have sufficed.
- Neglecting the squad-count check.
- Not exploiting stoppages as resynchronisation points.
- Annotating events with no written definition of the start instant.
- Entrusting event annotation to an annotator with no sporting competence.
- Working in image coordinates when the pitch supplies a reference.
- Ignoring the speed check as a switch detector.
- Using a broadcast camera for a deliverable requiring complete coverage.
- Renaming tracks manually rather than addressing their continuity.
- Costing without separating the positions line from the events line.
