jeremyipark/vision-demos

Fun real-world computer vision demos!

What it solves

This repository provides a collection of real-world computer vision demos that apply pose estimation and segmentation to analyze physical activities. It solves the problem of extracting quantitative data from video of human movement, such as athletic performance, dance synchronization, and exercise form analysis.

How it works

The project uses a variety of advanced vision models to process video clips. It leverages vitpose-plus-large for pose estimation across all demos, and sam3.1 (Segment Anything Model) for object segmentation. For specific tasks like deadlift form analysis, it integrates a multimodal LLM (gemma-4-26b-a4b-it) to provide a verdict on back posture.

Who it’s for

This is for developers and researchers interested in seeing how computer vision models can be applied to specific, high-precision movement analysis tasks like sports science, fitness tracking, and choreography analysis.

Highlights

  • Dance Sync: Computes a similarity metric to compare how well dancers are synchronized in the same choreography.
  • Rock Climbing: Segments bouldering holds and tracks the order of holds used, with a 3D version that uses iPhone LiDAR depth data to measure routes in meters.
  • Fitness Tracking: Includes tools to count reps for chin-ups and deadlifts, and measure running cadence and foot strike timing.
  • Deadlift Analysis: Classifies whether a back is straight or rounded during a deadlift rep with a probability score.

Related

  • Project
    yo-WASSUP/Good-GYM

    An AI fitness assistant that uses RTMPose for real-time pose detection and automatic exercise repetition counting via a webcam.

  • Project
    perfanalytics/pose2sim

    Pose2Sim is a free, open‑source Python toolkit that converts ordinary multi‑camera video into research‑grade 3‑D kinematics (OpenSim joint angles and marker trajectories). It works with any camera, supports multiple people, runs locally for privacy, and offers GPU‑accelerated pose estimation via RTMPose (or other models).

  • Project
    LibreYOLO/libreyolo

    An MIT-licensed computer vision library that provides a unified API for over twenty vision tasks, including detection, segmentation, and pose estimation, with integrated training and export tools.

  • Project
    AmmarkoV/SAM3DBody-cpp

    A standalone C++ inference engine for 3D human body pose estimation that regresses 3D mesh and pose parameters directly from images without Python dependencies.

  • Project
    fossasia/visdom

    Visdom is a Python‑driven web dashboard for live visualisation of plots, images, text and more. It organizes visualisations into persistent *environments* and draggable *windows*, supports callbacks, HTTPS/auth, and an async client, making it a lightweight tool for monitoring and sharing machine‑learning experiments.