VolETA++

Code for VolETA++: A Fully Automatic Depth-Free Radiance-Field Framework for Food 3D Reconstruction and Volume Estimation, A. AlMughrabi, R. Marques and P. Radeva, CVPR Workshops (3rd MetaFood Workshop), 2026, pp. 9482–9491.

This repository is a snapshot of the VolETA code base (release 0.0.3, March 2025), the version used for VolETA++. The code base was later extended for RASS (NeurIPS 2026); that version is at github.com/GCVCG/RASS.

Components in the VolETA repository. Keyframe selection (key_frame_extraction.sh, with blur filtering in blur_fft/ and near-duplicate removal in imgdp.py) and featuremetric pose estimation (pixelsfm.py) are not in this repository. Use them from github.com/GCVCG/VolETA-MetaFood.

Contents

• scripts/: pipeline stages (COLMAP, FoodMem masks, NeuS2 and Instant-NGP training, scaling, volume estimation).
• src/: pose conversion, checkerboard scaling (src/pose/find_scale.py), mesh clean-up, scaling and volume computation.
• docker/: Dockerfiles for COLMAP, PixSfM, Pre-NeRF 360 and the VolETA environment (see docker/README.md).

Unpublished experiments of the thesis

thesis-experiments/ holds the code and per-run data of the additional, unpublished results in Section 9.8 of the candidate's doctoral thesis: the variation of the NeuS2 volume between training runs, the effect of the training settings and masks of VolTex and VolETA++, and a re-run of the scale step. They are not peer-reviewed.

Settings used in the paper

• NeuS2: 15,000 iterations, marching-cubes resolution 512, and in each transforms.json "aabb_scale": 1, "scale": 0.15, "offset": [0.5, 0.5, 0.5] (scripts/run_neus2.sh uses these values).
• Mesh clean-up: isolated pieces below 5% of the mesh diagonal are removed.
• Scale: checkerboard squares of 0.012 m and a binarisation threshold of 240 for the corner detector (src/pose/find_scale.py). The threshold is sensitive: on some scenes the board is found in enough keyframes only at 210 to 220.

Acknowledgement of third-party code

src/pose/find_scale.py is adapted from the corner-matched scaling of the ININ-VIAUN team (HR-CMS in the MetaFood CVPR 2024 challenge report), published at github.com/BITyia/cvpr-metafood (scale_estimate/find_scale.py). Our changes add connected-component labelling and convex hulls to isolate the checkerboard before corner detection. The original repository states no licence; that file remains under its authors' terms and is not covered by the licence of this repository.

Licence

GNU General Public License v3.0 (see LICENSE), except for the third-party code below, which remains under the terms of its authors:

• src/pose/find_scale.py, adapted from HR-CMS (see above).
• thesis-experiments/code/find_scale-headless.patch, a change to that file, under the same terms.
• thesis-experiments/code/neus2-seed-headless.patch, a change to NeuS2, which keeps the NVIDIA Source Code License of NeuS2 (and the BSD licence of pybind11 for its pybind11 lines).
• src/colmap2nerf.py, from NVIDIA's instant-ngp (scripts/colmap2nerf.py), which keeps its NVIDIA copyright notice and the NVIDIA Source Code License. Helper functions taken from the same script (for example closest_point_2_lines) also appear in src/ARCORE_TO_NERF.py, src/ARCORE_TO_NERF_LOGMEAL.py and src/pose/pose_extraction.py.
• src/colmap/read_write_model.py and src/colmap/read_write_dense.py, from COLMAP (ETH Zurich and UNC Chapel Hill, BSD licence), which keep their copyright notices.

Citation

@inproceedings{almughrabi2026voletapp,
  author    = {AlMughrabi, Ahmad and Marques, Ricardo and Radeva, Petia},
  title     = {{VolETA++}: A Fully Automatic Depth-Free Radiance-Field Framework for Food {3D} Reconstruction and Volume Estimation},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)},
  pages     = {9482--9491},
  year      = {2026},
}