MMPose — OpenMMLab Pose Estimation Toolbox

Introduction

MMPose is a comprehensive pose estimation toolbox from the OpenMMLab ecosystem. It supports diverse tasks from human body keypoints to hand gesture recognition and animal pose tracking, all through a consistent modular API backed by PyTorch.

What MMPose Does

• Estimates 2D and 3D keypoints for human body, hands, face, and animals
• Implements 30+ methods including HRNet, RTMPose, and ViTPose
• Provides top-down and bottom-up pose estimation pipelines
• Supports whole-body pose estimation combining body, hand, and face
• Integrates with MMDetection for person detection before pose estimation

Architecture Overview

MMPose follows a top-down or bottom-up paradigm. Top-down first detects each person with a bounding box (via MMDetection), then estimates keypoints within each box. Bottom-up detects all keypoints simultaneously and groups them by person. Both approaches use configurable backbones, heads, and codec modules managed by MMEngine.

Self-Hosting & Configuration

• Install mmpose, mmengine, mmcv, and optionally mmdet via pip
• Download model checkpoints from the MMPose model zoo
• Use config files to select backbone, keypoint head, and dataset
• Set input resolution to balance speed and accuracy
• Deploy with MMDeploy for ONNX or TensorRT inference

Key Features

• RTMPose models achieve real-time performance at high accuracy
• Unified framework for body, hand, face, and animal keypoints
• Extensive model zoo with pre-trained weights on COCO, MPII, and more
• Modular codec system for keypoint encoding and decoding
• Built-in visualization with skeleton overlay on images and video

Comparison with Similar Tools

• MediaPipe — optimized for mobile and web but closed ecosystem; MMPose offers more research flexibility
• OpenPose — pioneered real-time pose but is slower; RTMPose in MMPose is faster and more accurate
• Detectron2 — supports keypoint detection but with fewer pose-specific methods
• AlphaPose — strong real-time performance but narrower scope than MMPose

FAQ

Q: Can MMPose track poses across video frames? A: MMPose handles per-frame estimation. Combine with a tracker like ByteTrack for temporal tracking.

Q: Does it support 3D pose estimation? A: Yes. MMPose includes 3D pose methods that lift 2D keypoints into 3D coordinates.

Q: What is RTMPose? A: RTMPose is a real-time pose estimation model in MMPose that achieves state-of-the-art speed-accuracy tradeoffs.

Q: Can I train on custom keypoint definitions? A: Yes. Define a custom dataset class with your keypoint schema and skeleton connectivity.

Sources

• https://github.com/open-mmlab/mmpose
• https://mmpose.readthedocs.io/