Fundamental Fencing Action Detection
Explore video-based identification of candidate épée actions, their classification confidence, and the chronological movement and distance evidence associated with each segment.
What action detection reports
FencingMonitor evaluates eligible épée video segments and presents candidate classifications for both fencers. The result is a structured computational assessment of observed movement, not a declaration of tactical intention or certainty.
The report connects relevant moments to their reference frames and observed distance, allowing users to review the sequence against the source video and the available skeleton visualization.
Eleven action categories
The classification vocabulary covers the following candidate actions. Names describe report categories; they are not automatic statements of refereeing priority or tactical correctness.
1. Attack on Preparation
An offensive action initiated by fencer A during opponent B’s preparation—such as an advance, blade action or feint—before B initiates an offensive action of their own. The classification does not depend on whether a touch is scored.
2. Counter-Attack
An offensive or offensive-defensive action executed by B during A’s ongoing offensive action, involving B’s own forward offensive body progression. In the FencingMonitor classification, this category is distinguished from an arm-led Stop-Hit.
3. Stop-Hit
A specific form of counter-attack in which B intervenes offensively with the weapon arm during A’s attack, without initiating forward offensive body progression. B may remain stationary or retreat defensively.
4. Parry-Riposte
A defensive-offensive sequence in which B makes a parry against A’s offensive action and then ripostes. The riposte may be immediate or delayed. Video-based classification may remain a candidate when blade contact cannot be directly confirmed.
5. Second Intention
A fencing sequence in which A’s initial offensive action is intended to provoke B’s response, followed by a second action exploiting that response. FencingMonitor identifies an observable sequence compatible with second intention, not the athlete’s unobservable mental intent.
6. Exit on Time
An offensive action initiated by B in anticipation of a possible offensive action from A, aiming to intercept or prevent its initiation or development. It does not require forward body progression, a successful touch, or even A’s eventual attack. The system can only assess an observable pattern compatible with this tactical concept.
7. Remise
A simple and immediate renewal of A’s prior offensive action, without withdrawing the weapon arm or returning to guard, after B has parried or evaded the first action by retreating and has not riposted immediately, or has used a delayed, indirect or compound riposte.
8. Reprise
A new offensive action launched immediately after returning to guard following an earlier unsuccessful offensive action. In FencingMonitor, it is distinguished from the specific forward-recovery-and-new-lunge pattern of Redoublement.
9. Redoublement
A renewed offensive action in which A recovers forward to guard after an initial offensive action, then performs another lunge. The weapon arm may remain partially extended during the transition. This is a FencingMonitor operational distinction informed by French fencing terminology.
10. Pressure Advance
A forward progression pattern in which A advances without initiating a full offensive action, reducing interpersonal distance or applying spatial pressure to B. The opponent may retreat or adjust position. This is a movement context rather than a single instantaneous attack.
11. Flèche
An offensive action in which the weapon-arm extension begins toward the target before forward projection of the body, with forward imbalance and alternating leg propulsion so the rear foot passes the front foot, without returning to guard before the action ends. Arm extension and body projection may partly overlap.
Classification confidence, not certainty
Candidate actions are presented with computational confidence scores expressing the system’s support for each classification within the analyzed segment. The leading candidate is not necessarily the only plausible interpretation.
These scores are model-derived indicators, not guarantees of correctness or independently calibrated statistical probabilities. Ambiguous and incomplete visual evidence can affect candidate ranking.
Sequence, frames and distance
Frame-referenced sequence
Important phases and transitions are linked to identifiable video frames or frame intervals so that observations can be checked against the recording.
Distance evolution
The movement sequence is read alongside observed changes in the distance between the two fencers; a distance measurement alone does not establish timing, intention, or a specific action.
Biomechanical context
Movement observations and measured or derived biomechanical indicators provide context for classification without independently proving its tactical meaning.
Scientific and Research Framework
Research scope
The framework investigates how observable kinematics, interpersonal distance, and chronological movement organization relate to candidate fencing actions in real video footage.
Measurement foundation
Markerless video pose estimation supplies movement observations. FencingMonitor’s computational core produces measurements, biomechanical metrics, derived indicators and classification outputs.
Analytical transparency
Candidate ranking, explicit confidence scores, frame references and contextual observations help make outputs inspectable and suitable for further expert review.
Scientific boundaries
This is a video-based analytical instrument, not a claim of laboratory-grade accuracy or independent scientific validation. Camera position, occlusion, image quality and modeling assumptions can limit reliability.
Generative AI is not the source of measurements, calculations or action classifications. If used, an external AI layer is optional and downstream, for language support, synthesis of existing results and exploration of clearly labeled hypotheses.
Video and interpretation requirements
Eligible clips must show both fencers, last no more than 15 seconds, be at least 720p and 30 fps, and use MP4 or MOV. Blur, occlusion, poor framing or camera motion may produce partial or unreliable output. Consult the Video Recording Guide.
Interpretation boundaries
FencingMonitor does not issue official refereeing decisions, infer the athlete’s intent as a measured fact, prescribe individual training, or guarantee classification accuracy or improved performance. Expert review remains essential, particularly where evidence is ambiguous.
Availability and ongoing refinement
Action detection is presented as a FencingMonitor capability. Classification models and their evaluation may continue to be refined; users should interpret results with the reported confidence and evidence limitations.