PoseEstimationForMobile  by edvardHua

Pose estimation for mobile devices

created 7 years ago
1,015 stars

Top 37.5% on sourcepulse

GitHubView on GitHub
Project Summary

This repository provides real-time single-person pose estimation models (CPM and Hourglass) optimized for mobile devices using TensorFlow and inverted residuals (MobileNet V2). It targets mobile developers and researchers seeking efficient pose estimation solutions, offering significant speed improvements on mobile hardware.

How It Works

The project implements Convolutional Pose Machines (CPM) and Hourglass network architectures, enhanced with MobileNet V2's inverted residual blocks. This design choice reduces model complexity and computational cost (FLOPs), enabling real-time inference on mobile CPUs and GPUs. The models are trained on the AI Challenger dataset and can be converted to formats like TFLite and CoreML for deployment.

Quick Start & Requirements

  • Training: Requires Python 3, TensorFlow >= 1.4, Mace, and the AI Challenger dataset. Training can be performed via Docker or directly on macOS.
  • Inference: Pre-trained models (frozen graphs, TFLite, CoreML) are available. Android demo uses Mace or TFLite; iOS demo uses CoreML.
  • Links:
    • Android APKs: [PoseEstimation-Mace.apk](Using Mace (Support GPU)), [PoseEstimation-TFlite.apk](Using TFlite (Only CPU))
    • Mace Documentation: [Mace documentation]([7] Mace documentation)
    • TensorFlow Lite Guide: [Devlope guide of TensorFlow Lite]([6] Devlope guide of TensorFlow Lite)

Highlighted Details

  • Achieves ~60 FPS on Snapdragon 845 and ~30 FPS on iPhone X for CPM/Hourglass models.
  • Offers model conversion tools for TFLite (CPU/GPU) and CoreML (iOS).
  • Includes full Android and iOS demo source code.
  • Benchmarks provided for various mobile chipsets (Snapdragon, Exynos) and iPhones.

Maintenance & Community

  • Actively welcomes Issues and Pull Requests.
  • References several key research papers and related repositories.

Licensing & Compatibility

  • Licensed under the Apache License 2.0.
  • Permits commercial use and integration into closed-source applications.

Limitations & Caveats

The repository is presented as a baseline, with significant room for improvement in model architecture and dataset utilization. Mace framework build is not supported on Windows.

Health Check
Last commit

2 years ago

Responsiveness

Inactive

Pull Requests (30d)
0
Issues (30d)
0
Star History
5 stars in the last 90 days

Explore Similar Projects

Feedback? Help us improve.