SURREAL Mesh Export Fork

This repository is a focused fork of SURREAL for exporting posed human walking meshes. The intended use is robotics data augmentation: generate one 3D human mesh per walking animation frame, then place those meshes into a separate scene or simulator.

The primary workflow is mesh-only. It does not require the original RGB/depth rendering pipeline, LSUN backgrounds, ffmpeg, OpenEXR, Torch training code, or pretrained SURREAL models.

What This Fork Produces

The export path in datageneration/export_obj_frames.py loads SURREAL SMPL motion data and writes:

• one OBJ file per animation frame, for example frame_000000.obj
• metadata.json describing the source frames, axes, units, topology, and texture choices
• optional material.mtl and copied clothing texture files

Meshes are exported in meters, pelvis-centered in XY, grounded per frame, with target axes +X forward, +Y left, and +Z up.

Essential Setup

Only the assets required for mesh export need to be installed.

1. SMPL FBX Models

Download the licensed SMPL for MAYA models from the SMPL website after accepting their license terms. Place these files under datageneration/smpl_data/:

datageneration/smpl_data/
  basicModel_f_lbs_10_207_0_v1.0.2.fbx
  basicModel_m_lbs_10_207_0_v1.0.2.fbx

2. SURREAL SMPL Motion Data

Download the SURREAL SMPL data after accepting the SURREAL license terms. The mesh export script needs:

datageneration/smpl_data/
  smpl_data.npz

This file contains the CMU MoCap-derived SMPL pose and translation sequences used by the exporter.

3. Optional Clothing Textures

If textured OBJ output is desired, also place the SURREAL texture data under:

datageneration/smpl_data/textures/

Use --no-texture when exporting geometry only.

4. Blender

Install Blender. The helper script uses BLENDER_PATH if it is set; otherwise it falls back to the local path currently encoded in datageneration/export_obj_frames.sh.

Example:

export BLENDER_PATH=/path/to/blender

Export Meshes

See datageneration/README.md for the full export guide.

Minimal example:

cd datageneration
./export_obj_frames.sh --idx 02_02 --gender female --frames 30 --stepsize 4 --out ../outputs/obj_frames

When choosing a walking sequence, start with datageneration/misc/walking_sequences.txt. It lists the walking sequences that have been checked for this fork. datageneration/misc/sequence_idx_map lists broader sequence names available in the SURREAL motion data.

Original SURREAL Project

This fork is based on:

Gul Varol, Javier Romero, Xavier Martin, Naureen Mahmood, Michael J. Black, Ivan Laptev, and Cordelia Schmid, Learning from Synthetic Humans, CVPR 2017.

Original resources:

• Project page
• Paper
• Original repository

Citation

If you use this code or the SURREAL assets, please cite the original work:

@INPROCEEDINGS{varol17_surreal,
  title     = {Learning from Synthetic Humans},
  author    = {Varol, G{\"u}l and Romero, Javier and Martin, Xavier and Mahmood, Naureen and Black, Michael J. and Laptev, Ivan and Schmid, Cordelia},
  booktitle = {CVPR},
  year      = {2017}
}

License

Check the repository license in LICENSE.md and the SURREAL/SMPL asset license terms before downloading, using, or redistributing code, models, or data.