Implementation:Datajuicer Data juicer ImageMMPoseMapper

Knowledge Sources	Datajuicer_Data_juicer
Domains	Data_Processing, Mapping
Last Updated	2026-02-14 16:00 GMT

Overview

Concrete tool for performing human keypoint detection (pose estimation) on images using MMPose provided by Data-Juicer.

Description

ImageMMPoseMapper is a mapper operator that performs human keypoint detection inference using MMPose models deployed via MMDeploy. It requires three essential components: a deployment config (deploy_cfg), a model config (model_cfg), and model weight files (model_files). The operator automatically installs required packages (openmim, mmpose, mmdet) if not present. It processes each image to extract keypoints, keypoint scores, bounding boxes, and bbox scores, storing results in sample metadata under a configurable pose key. Supports optional visualization output. Requires CUDA acceleration.

Usage

Use when you need to extract skeletal keypoint data from images for human-centric computer vision datasets and body-aware data filtering.

Code Reference

Source Location

• Repository: Datajuicer_Data_juicer
• File: data_juicer/ops/mapper/image_mmpose_mapper.py

Signature

@OPERATORS.register_module("image_mmpose_mapper")
class ImageMMPoseMapper(Mapper):
    def __init__(self,
                 deploy_cfg: str = None,
                 model_cfg: str = None,
                 model_files: Optional[Union[str, Sequence[str]]] = None,
                 pose_key: str = MetaKeys.pose_info,
                 visualization_dir: str = None,
                 *args, **kwargs):

Import

from data_juicer.ops.mapper.image_mmpose_mapper import ImageMMPoseMapper

I/O Contract

Inputs

Name	Type	Required	Description
deploy_cfg	str	Yes	MMPose deployment config file path
model_cfg	str	Yes	MMPose model config file path
model_files	Optional[Union[str, Sequence[str]]]	Yes	Path to model weight files
pose_key	str	No	Key to store pose information in metadata, defaults to MetaKeys.pose_info
visualization_dir	str	No	Directory to save visualization results; if None, no visualization is saved

Outputs

Name	Type	Description
samples	Dict	Transformed samples with pose keypoint data stored in meta field

Usage Examples

process:
  - image_mmpose_mapper:
      deploy_cfg: "/path/to/deploy_cfg.py"
      model_cfg: "/path/to/model_cfg.py"
      model_files: "/path/to/model.pth"

Related Pages

• Environment:Datajuicer_Data_juicer_Python_Runtime_Environment