Skip to the work
← All work

Computer vision + fight analysis

Cageframe & Prometheus

Teaching a machine to read a fight.

Prototype — runs on my machinecageframe.com →
PythonComputer visionPose estimationVision-language modelsBlender

The question it started from

Can software watch a round of MMA and tell you what actually happened — not just that someone moved, but which technique, in what sequence, and what decision it implies?

The problem

Fight analysis lives in the heads of coaches and commentators. The raw material is enormous and completely unreadable by software: to a generic vision model, a takedown, a level change and a check hook are the same blur of two bodies.

And the hardest part is not detection. It is that a fight is a sequence — a technique only means something next to the one before it.

The idea

Build the chain that turns footage into structured fight knowledge, one layer at a time: find the fighters, lift them into pose, name the technique, then read sequences and decisions on top of that representation.

Keep every layer inspectable. If the model is wrong about a technique, I want to see which stage lied.

What I contributed

I designed and built the pipeline end to end: the extraction loop over video, pose and visibility handling, the prompting and scoring layer that names techniques, and the pages that make a fight readable.

I also did the visual side — skeleton overlays, annotated cards and 3D remakes of specific exchanges in Blender, used both to debug the model and to explain a moment to a human.

The pipeline
  1. 01

    Footage

    A video and a time window — one exchange, not a whole fight.

  2. 02

    Detect & track

    Find both fighters and keep their identity across frames.

  3. 03

    Pose

    Lift each fighter into joints, with a visibility score per joint.

  4. 04

    Technique

    Name what is happening, asking only questions the frame can support.

  5. 05

    Sequence

    Chain techniques into an exchange that has a beginning and an outcome.

  6. 06

    Read

    Say what the decision was — and show the evidence for it.

What exists today

  • A repeatable loop that takes a video plus a time window and produces an annotated page for that exchange.
  • Pose extraction with per-joint visibility, so the model knows which parts of the body it cannot actually see.
  • A technique-naming layer where candidate prompts are filtered by how much of the body is visible, and overlapping answers are reconciled rather than averaged.
  • 3D reconstructions of individual exchanges, rebuilt from the footage for explanation.
  • A public page at cageframe.com — today a waitlist, not the product. The pipeline itself runs on my own machine, not on a server.

What is left to build

  • Sequence reading is still the frontier — naming single techniques is further along than reading how they chain.
  • Ground positions remain the weak spot: that is where monocular footage hides the most.
  • Rights and licensing decide what footage this can ever run on at scale. That is a product constraint, not an engineering one.

What it taught me

Monocular video hides more than it shows. Depth-based visibility stops being trustworthy the moment the fight goes to the ground, so the pipeline has to reason explicitly about what is occluded instead of pretending it sees everything.

Filtering the question by what is visible beat reaching for a bigger model. Narrowing the candidate set to techniques the frame could actually support was the change that moved accuracy.

Rebuilding an exchange in 3D is the fastest way to find out that you have misunderstood it.

Next

Signl