What is densepose?

facebookresearch/densepose — explained in plain English

Analysis updated 2026-06-24

★ 7,230Jupyter NotebookAudience · researcherComplexity · 4/5Setup · hard

Mindmap

mindmap
  root((DensePose))
    What it does
      Pixels to 3D body
      Full body at once
      Not just joints
    Dataset
      DensePose-COCO
      SMPL body template
      Interactive notebooks
    Tech Stack
      Python
      Caffe2
      Detectron2
    Status
      Archived 2018 paper
      Moved to Detectron2
      Reference only

mindmap root((DensePose)) What it does Pixels to 3D body Full body at once Not just joints Dataset DensePose-COCO SMPL body template Interactive notebooks Tech Stack Python Caffe2 Detectron2 Status Archived 2018 paper Moved to Detectron2 Reference only

Click or tap to explore — scroll the page freely

What do people build with it?

USE CASE 1

Map body pixels in a photo to 3D surface coordinates for animation or augmented reality.

USE CASE 2

Browse and visualize DensePose-COCO annotations overlaid on a 3D SMPL body template.

USE CASE 3

Reproduce the original 2018 CVPR DensePose-RCNN results for academic research.

What is it built with?

PythonJupyter NotebookCaffe2Detectron2

How does it compare?

	facebookresearch/densepose	harvardnlp/annotated-transformer	rasbt/python-machine-learning-book-2nd-edition
Stars	7,230	7,248	7,201
Language	Jupyter Notebook	Jupyter Notebook	Jupyter Notebook
Setup difficulty	hard	easy	moderate
Complexity	4/5	2/5	2/5
Audience	researcher	researcher	researcher

Figures from each repo's GitHub metadata at analysis time.

How do you get it running?

Difficulty · hard Time to first run · 1day+

This repo is archived, the maintained version lives in Detectron2 and requires a matching Caffe2/PyTorch environment.

So what is it?

DensePose is a research project from Facebook that takes a regular photo of a person and figures out how every visible pixel on their body maps to a 3D model of the human form. In plain terms: you give it an image of someone standing, sitting, or moving, and the system can tell you exactly which part of a 3D body surface each pixel corresponds to, covering the whole visible body at once rather than just tracking joints or outlines. The code here was used to train and test the original DensePose-RCNN system, which was published at a computer vision conference in 2018. It also includes interactive notebooks that let you browse the DensePose-COCO dataset, a large collection of images with body-surface annotations, and see how those annotations look when placed on a 3D body template called SMPL. This repository is no longer actively maintained. The project has moved into Detectron2, a newer Facebook research framework, where it continues to receive updates and new model versions. If you want to use DensePose today, the Detectron2 version is the right starting point. This older repository is preserved mainly as a reference for the original 2018 paper.

Copy-paste prompts

Prompt 1

I have an image of a person. Using DensePose, write Python code to predict which 3D body surface region each visible pixel maps to and display the result as a colored overlay.

Prompt 2

Show me how to load a DensePose-COCO annotation file and visualize the body surface UV coordinates on the SMPL template using the provided Jupyter notebooks.

Prompt 3

I want to run DensePose inference on my own photos. Point me to the Detectron2 version of DensePose and show me the minimal setup steps to get predictions.

Prompt 4

Explain the difference between DensePose's pixel-to-surface mapping and standard human pose estimation that only predicts joints.

Frequently asked questions