How open-mmlab/mmpose works
MMPose is a PyTorch toolbox for 2D and 3D pose estimation across human, animal, face, hand, whole-body, and fashion tasks. A typical inference flow takes an image or video through an inferencer, prepares pose inputs, runs a configured model, decodes keypoints, and returns predictions that can be visualized. The repository also provides dataset and evaluation components and configurable model families. Edges reflect explicit imports in indexed source or documented capabilities; unsampled implementation details are not inferred.
Main entry points: base_mmpose_inferencer.py Pose Inferencer · pose3d_inferencer.py 3D Inferencer · configs Task Configurations · inference.py Inference API
Components 24
Inference APIs
- Pose Inferencer
mmpose/apis/inferencers/base_mmpose_inferencer.py - 2D Inferencer
mmpose/apis/inferencers/pose2d_inferencer.py - 3D Inferencer
mmpose/apis/inferencers/pose3d_inferencer.py - Inference API
mmpose/apis/inference.py - 3D Inference
mmpose/apis/inference_3d.py
Pose Data
- Datasets
mmpose/datasets/datasets - Data Transforms
mmpose/datasets/transforms/common_transforms.py - Pose Structures
mmpose/structures/pose_data_sample.py
Model Components
- Task Configurations
configs - Component Registry
mmpose/registry.py - Feature Backbones
mmpose/models/backbones - Pose Heads
mmpose/models/heads - Keypoint Codecs
mmpose/codecs - Associative Embedding
mmpose/codecs/associative_embedding.py
Evaluation & Output
- Keypoint Metrics
mmpose/evaluation/metrics/keypoint_2d_metrics.py - Evaluation Functions
mmpose/evaluation/functional/keypoint_eval.py - Pose Visualizer
mmpose/visualization/local_visualizer.py
Applications
- RTMPose Demo
projects/rtmpose/app.py - Dance Demo
projects/just_dance/app.py - Video Processor
projects/just_dance/process_video.py - TorchServe Handler
tools/torchserve/mmpose_handler.py
Other components
- User
- Image or Video
- Pose Predictions
Connections 22
- User → Pose Inferencer: requests inference
README.md - Image or Video → Pose Inferencer: provides input
mmpose/apis/inferencers/base_mmpose_inferencer.py - Pose Inferencer → Inference API: calls
mmpose/apis/inferencers/base_mmpose_inferencer.py - 2D Inferencer → Pose Inferencer: extends
mmpose/apis/inferencers/pose2d_inferencer.py - 3D Inferencer → Pose Inferencer: extends
mmpose/apis/inferencers/pose3d_inferencer.py - 3D Inferencer → 2D Inferencer: uses
mmpose/apis/inferencers/pose3d_inferencer.py - 3D Inferencer → 3D Inference: uses
mmpose/apis/inferencers/pose3d_inferencer.py - Inference API → Pose Structures: returns samples
mmpose/apis/inference.py - Task Configurations → Datasets: configures
README.md - Task Configurations → Feature Backbones: configures
README.md - Task Configurations → Pose Heads: configures
README.md - Data Transforms → Keypoint Codecs: uses
mmpose/datasets/transforms/common_transforms.py - Associative Embedding → Keypoint Codecs: extends
mmpose/codecs/associative_embedding.py - Keypoint Codecs → Component Registry: registers codecs
mmpose/codecs/associative_embedding.py - Pose Heads → Component Registry: registers models
- Feature Backbones → Component Registry: registers models
mmpose/models/backbones/litehrnet.py - Keypoint Metrics → Evaluation Functions: calls
mmpose/evaluation/metrics/keypoint_2d_metrics.py - Pose Inferencer → Pose Visualizer: visualizes
mmpose/apis/inferencers/base_mmpose_inferencer.py - Pose Inferencer → Pose Predictions: produces
mmpose/apis/inferencers/base_mmpose_inferencer.py - RTMPose Demo → Pose Inferencer: calls
projects/rtmpose/app.py - Dance Demo → Video Processor: calls
projects/just_dance/app.py - TorchServe Handler → Inference API: uses API
tools/torchserve/mmpose_handler.py