SimonPose

SimonPose is a research prototype exploring how to improve monocular human pose estimation by fitting a parametric body model (SMPL) to DensePose v2 / CSE outputs.

DensePose provides dense 2D-to-body correspondences (per-pixel embeddings). SimonPose uses those correspondences to optimize SMPL parameters and produce a globally consistent 3D human mesh.

This repository accompanies the write-up for maturita exam in documentation/simon-bravek-mp-2026.pdf.

Requirements (minimal)

Tested stack (recommended):

• Python 3.11
• Linux + NVIDIA GPU
• PyTorch 2.4.x + CUDA (tested on CUDA 12.4 with cu121 PyTorch wheels)

It may run on CPU, but DensePose + fitting will be very slow.

Project status

• Research code (not a polished library); scripts are the primary interface.
• projects/euclidean_fitter.py: usable end-to-end experiment script (SMPL fitting driven by a 2D euclidean loss).
• projects/precission_fitter.py: work-in-progress analysis prototype; computes/visualizes DensePose correspondence stability, but does not run a full SMPL fitting loop.

Projects

• projects.euclidean_fitter (ready): fits SMPL to DensePose CSE correspondences and saves per-iteration visualizations + loss plots.
• projects.precission_fitter (WIP): analyzes DensePose-to-SMPL correspondence stability (per-vertex dispersion) and visualizes it; no fitting/optimization yet.

The report also discusses an embedding-based fitter prototype; it was evaluated as a dead end and is not maintained as a runnable project in this repository.

Inputs / outputs

• Input: RGB images (COCO 2014 val by default) + DensePose CSE weights + SMPL model file.
• Output: visualizations and diagnostics in output/ (loss plots, overlays of the fitted mesh on the image).

Quickstart

You run everything from the repo root.

Full documentation index: docs/README.md

1. Setup environment and dependencies:

• Local machine: docs/setup-local.md
• CTU cluster: docs/setup-ctu.md

2. Download datasets and model files:

• docs/data.md

3. Run the working project:

python -m projects.euclidean_fitter -n 1 -i 200

Expected output:

• A new run directory output/euclidean_fitter_0/ (or the next available suffix).
• Per-image overview PNGs like output/euclidean_fitter_0/<image_id>_<person_index>_overview.png.
• Aggregate plots: output/euclidean_fitter_0/loss_histogram.png and output/euclidean_fitter_0/loss_remaining_histogram.png.

Running experiments

• How to run each script, what it produces, and what is WIP: docs/running.md

Repository layout

• projects/: runnable experiment scripts.
• common/: shared utilities (projection, loss functions, DensePose overrides, optional depth experiments).
• config.py: central paths; assumes this directory structure:
  • data/ (ignored by git): datasets and annotations
  • models/ (ignored by git): DensePose + SMPL model files
  • external/ (ignored by git): third-party code such as Detectron2
  • output/ (ignored by git): generated results

If you want a different layout, edit config.py.

Configuration details: docs/configuration.md

Troubleshooting

• Common install/runtime issues (Detectron2 builds, CUDA mismatch, missing models/data): docs/troubleshooting.md

License

The code in this repository is licensed under Apache 2.0 (see LICENSE).

Note: datasets and model files (COCO, DensePose weights, SMPL/SMPL-X) are not redistributed and have their own licenses/terms.

Credits

This project relies on:

• Detectron2 + DensePose (Apache 2.0)
• COCO API (Apache 2.0)
• SMPL/SMPL-X model files (separate license/terms from MPI/MPG; see docs/data.md)

Citation

If you use this code in academic work, please cite the accompanying report:

@misc{bravek2026simonpose,
  title        = {SimonPose},
  author       = {Šimon Brávek},
  year         = {2026},
  institution  = {FEL CTU (fel.cvut.cz) and Johannes Kepler Grammar School (gjk.cz)},
  howpublished = {[simon-bravek-mp-2026.pdf](https://github.com/simonbravek/SimonPose/blob/main/documentation/simon-bravek-mp-2026.pdf)}
}