Practical overview
When it fits
Use it when you need implementation code for camera-distance-aware multi-person pose estimation, or a documented demo that accepts an image, pretrained snapshot and configured bounding boxes.
Who it is for
- Computer-vision researchers studying 3D human pose estimation.
- Developers building or exploring multi-person pose-estimation prototypes with Python and PyTorch.
Documented uses
- Estimate root-joint depth from an RGB image using the documented demo.
- Explore camera-distance-aware 3D pose estimation.
- Prepare supported pose datasets or COCO-format annotations for the repository's training and testing workflows.
- Visualize pose-estimation results.
Requirements & limits
The repository is public and MIT-licensed.
- The demo requires a separately obtained pretrained RootNet snapshot and a configured bounding-box list.
- The README lists PyTorch, CUDA, cuDNN, Anaconda and the COCO API as dependencies.
- The documented development setup is Python 3.6.5 with Anaconda 3; the reported test environment uses Ubuntu 16.04, CUDA 9.0 and cuDNN 7.1.
- Dataset-specific bbox_real units must be set correctly; the README gives Human3.6M and 3DPW examples.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Run the documented RootNet demo
Use the repository's demo with an image and pretrained RootNet snapshot to generate a root-estimation visualization and depth output.
Before you start
- The repository's demo code and its listed dependencies: PyTorch, CUDA, cuDNN, Anaconda and the COCO API.
- A pretrained RootNet snapshot and an input image named input.jpg in the demo folder.
- A bounding-box list configured in demo.py.
Steps
Place input.jpg and the pretrained snapshot in the demo folder.
Set bbox_list in demo.py.
From the demo folder, run the command below to use GPU 0.
python demo.py --gpu 0 --test_epoch 18
Expected outcome
The README describes output_root_2d.jpg and printed root-joint depths.
Guide limitations
- The documented development and testing environment is Ubuntu 16.04, CUDA 9.0, cuDNN 7.1, Python 3.6.5 with Anaconda 3, and two NVIDIA 1080Ti GPUs.
- The demo requires a pretrained snapshot, obtained separately from the repository.
Prepare a dataset for the repository's training or testing workflow
Organize dataset images and annotations in the repository's described structure and use the required annotation format.
Before you start
- Dataset images and annotations arranged under the documented data directory structure.
- Annotations in MS COCO format; custom datasets must be converted to that format.
- A dataset-appropriate bbox_real setting in main/config.py.
Steps
Create the dataset directories under data following the README's examples.
Place the relevant images and annotations in those directories.
Convert custom-dataset annotations to MS COCO format.
Set bbox_real in main/config.py using the units of the dataset.
Expected outcome
The dataset is organized in the format and directory layout the README specifies for the repository's training or testing workflows.
Guide limitations
- The README gives different bbox_real units for Human3.6M and 3DPW; the setting must match the dataset.
- This guide covers data preparation, not execution of training or testing.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 15 October 2026
- Original website title
- mks0601/3DMPPE_ROOTNET_RELEASE · Official homepage title · retained in its source language