What is captured from each video, how moves are labeled, and what the data will and will not be used to claim.
Sports like running and football have decades of motion capture research. Climbing has very little, and what exists rarely records what the climber was attempting or what they felt while doing it.
Dynalytix is building a labeled dataset of climbing movement. Every frame carries pose landmarks, joint angles, velocities, and center of mass, joined to a structured description of the move and any sensations the climber tagged.
The data has two layers:
Kinematics from pose estimation: where each joint is, how fast it is moving, and how the center of mass travels through the move.
What the wall offered, what they chose to do, how it turned out, and what it felt like, tagged on the frames where it happened.
Computed from the landmarks on both sides of the body, for every frame.
Full body tracking with x, y, z coordinates and visibility confidence.
Center of mass, individual landmarks, wrist/hip velocity vectors.
The export is one table: frame number, timestamp, pose features, and every label that applies to that frame.
What the wall affords. Wall angle (slab, vertical, gentle overhang, steep) and three holds per move: start left, start right, and the end hold the reaching hand goes to. Each hold gets a type (edge, gaston, side-pull, undercling), quality descriptors, and a box drawn on the video.
What the climber chose. An approach (static, dynamic, or coordination), any number of move tags (bump, mantle, balance, heel hook, toe hook, no feet, technical, tension, and others), and a size (small, medium, large). A 1 to 5 form quality score with anchored descriptions, and 0 to 10 perceived effort.
What happened. Success or fall, whether the hold was reached and controlled, and the labeler's confidence. Sensations are tagged on the frame where they occur: 9 types (sharp pain, dull pain, audible pop, unstable, stretch, weak, strong, pumped, fatigue), left or right, one of 16 body locations, and a 0 to 10 level.
Frame-accurate start and end markers for each move, set by the labeler while scrubbing the video.
Videos are stored privately. Uploaded videos, pose data, and labels are scoped to the account that created them, and access is enforced in the database. Videos are kept so pose can be re-extracted later with better models. Any public dataset release will be de-identified pose and label data, not video.
The work is planned as two papers, kept separate on purpose.
First, the dataset. The collection pipeline, the labeling schema, and the labeled data itself, including how consistently different labelers apply the schema to the same video. This is the near-term goal, and it is a contribution on its own: there is very little climbing movement data with the climber's own account attached.
Later, models. Only once the dataset is large enough to support real performance claims. No model has been trained yet, and nothing on this site should be read as a promise of injury prediction or personalized feedback.
Original videos are retained privately so the pose data can be re-extracted with better models in the future without asking anyone to climb again.
The pilot is starting with a small group. If you climb and would like to label video, or you work on climbing research and want to follow along, sign up for updates and say so in your reply.
Start contributing data