EvoSkeleton provides the official implementation for a CVPR 2020 paper on monocular 3D human pose estimation. It tackles inferring 3D poses from single images using a cascaded deep learning architecture and evolutionary training data synthesis. This project offers researchers advanced techniques for state-of-the-art 3D pose estimation.

How It Works

The system employs a cascaded 2D-to-3D lifting strategy, refining 2D keypoints into 3D skeletons. A key innovation is using an evolutionary algorithm for novel training data discovery, enhancing model robustness. It also features a high-resolution heatmap regression model for accurate 2D pose estimation, foundational for 3D lifting.

Quick Start & Requirements

• Setup requires a Python 3.6 environment, ideally using Anaconda and the provided spec-list.txt.
• Key dependencies: Python 3.6, Numpy 1.16, PyTorch 1.0.1, CUDA 9.
• Pre-trained models and usage instructions are available via linked sub-pages.
• Official Paper: [Link to Paper]
• Presentation Video: [Link to Video]

Highlighted Details

• Achieves superior Human3.6M performance, outperforming Martinez et al. (ICCV'17) on key metrics (e.g., Protocol #1 Avg. 49.7 vs 62.9).
• Improves 2D pose estimation accuracy (4.4 pixels error) over CPN (CVPR'18) on Human3.6M.
• Introduces the novel "Unconstrained 3D Human Pose in the Wild (U3DPW)" dataset via an interactive annotation tool.
• The evolutionary training data methodology offers a unique strategy for boosting model generalization.

Maintenance & Community

• Released v1.0 on April 8, 2021, enhancing pre-trained model support.
• Community interaction is via GitHub "discussions" (Q&A) and "issues" (technical problems). No direct links to community platforms are specified.

Licensing & Compatibility

• Repository code is MIT licensed.
• Use of third-party datasets (Human 3.6M) and tools (SMPLify) is subject to their licenses, potentially restricting commercial application.

Limitations & Caveats

• Reliance on outdated dependencies (Python 3.6, PyTorch 1.0.1) hinders integration with modern stacks.
• Commercial usage is constrained by third-party component licensing.
• No information on current maintenance or future development beyond the 2021 release is provided.