Computer vision - the field of AI that enables machines to interpret visual information - has been transforming professional sports for over a decade. Soccer clubs use it for player tracking. NBA teams use it to analyze shot mechanics. Baseball has used it to analyze pitcher release points since the mid-2010s.
Wrestling analytics lagged because of one fundamental challenge: the sport is dense, close-contact, and chaotic. The bodies overlap. The positions change in fractions of seconds. The action is continuous.
We built DUCKEYE™ to solve that problem specifically.
What Computer Vision Is
At its core, computer vision is teaching machines to extract structured information from images and video. The same way your brain looks at a wrestling match and sees "he's on top," a computer vision model can be trained to make the same inference - at scale, without fatigue, frame by frame.
The enabling breakthrough was the convolutional neural network (CNN), which can learn spatial patterns directly from image data rather than relying on hand-coded features.
How Wrestling Pose Estimation Works
The core technical challenge in wrestling analytics is pose estimation: determining the body position of each athlete in every frame of a video.
DUCKEYE™ uses a two-stage approach:
Stage 1: Person detection A detection model identifies each wrestler's bounding box in each frame. This needs to be robust to occlusion (bodies overlapping), motion blur, and variable camera angles.
Stage 2: Keypoint estimation For each detected person, a pose estimation model identifies 17-25 body keypoints - shoulders, elbows, hips, knees, etc. These keypoints define the skeleton of the athlete.
From the skeleton configuration, a classification model determines the positional state: neutral (standing), top, bottom, or transition.
The Training Data Challenge
The quality of a computer vision model is determined by the quality and quantity of training data. For wrestling, that meant:
- Footage across a wide range of body types, weight classes, and skill levels
- Multiple camera angles and lighting conditions
- Both men's and women's wrestling
- High school, collegiate, and international competition
DUCKEYE™'s model was trained on 10,000+ hours of footage with human-labeled positional annotations. That annotation process - frame by frame, match by match - is what makes the model generalizable.
When we say 85%+ accuracy, we mean across this diverse distribution. Not cherry-picked footage. Not a single camera setup. Diverse data.
Processing Pipeline
When you upload a match to DUCKEYE™, here's what happens:
1. Video ingestion: The file is received and validated. Format conversion if needed.
2. Frame extraction: The video is sampled at a fixed frame rate (typically 5-10 fps is sufficient for positional analysis; position states don't change faster than this).
3. Person detection: Each frame goes through the detection model. Two bounding boxes, one per wrestler.
4. Pose estimation: Keypoints are estimated for each detected person.
5. Position classification: The keypoint configuration maps to one of four position states.
6. Temporal smoothing: Raw frame-by-frame classifications are smoothed to remove noise and correctly handle ambiguous transition frames.
7. Analytics generation: The sequence of position states generates the timeline, percentages, and downstream analytics.
Total processing time: under 30 minutes for a full 7-minute match at standard quality.
Why Wrestling Is a Hard Computer Vision Problem
Wrestling is harder than many sports for a few specific reasons:
Occlusion: In grappling sports, athletes are physically connected. A wrestler on top partially obscures the wrestler on bottom. The model has to infer keypoints it can't directly observe.
Similar appearance: Two wrestlers in matching singlets look similar to a naive detector. Tracking individual athletes through a match requires maintaining identity across frames.
Variability: Wrestlers of different weight classes look dramatically different. A 125-pound neutral wrestler and a 285-pound neutral wrestler have very different spatial configurations. The model needs to be robust to this.
These are real challenges we've worked on. The 85%+ accuracy figure represents our current state of the art on diverse footage.
What's Next
We're actively working on:
- Live inference: Processing video in real-time rather than batch mode (planned for Elite tier)
- Individual technique classification: Beyond position states, identifying specific techniques (single leg, double leg, etc.)
- Predictive modeling: Using temporal patterns to predict likely next moves - the true frontier of competitive intelligence
The model improves with every match analyzed. Every new upload contributes to a better understanding of the variation in elite wrestling.
*Questions about our technical approach? Reach out at engineering@duckeyeanalytics.com*
Like this article? Try DUCKEYE™ free for 14 days.
No credit card required. Full access to all features.
Start Free Trial