# RefSeg-CA — R3 research release

Evaluating and repairing command compliance in generalized referring segmentation.

The release contains the SFAP parser/projection, benchmark generator, inference/evaluation scripts, exact-score replay, Slurm examples, editable TikZ/PGFPlots sources, and a current manuscript. Human validation has **not been conducted**. No acceptance or publication claim is made.

## Data and exact metric replay

Data repository: https://huggingface.co/datasets/Ethosoft/RefSeg-CA

From this directory, with Python 3.9 or newer:

```bash
python code/download_release_data.py
python code/verify_portable_results.py
```

The lightweight download reconstructs 30 primary table rows and 22 reviewer-requested threshold configurations. It requires no TRUBA connection or model weights. These are derived scores; reproducing inference pixels requires upstream models and inputs.

To obtain all 1,200 procedural scenes, both rendering styles, visible-instance maps, and command masks:

```bash
python code/download_release_data.py --synthetic
```

The default data revision is the immutable release tag `r3-data-v1`. Natural images and original gRefCOCO expression text must be obtained from the pinned upstream release; model checkpoint hashes and preparation scripts are under protocol/ and legacy_project/code/.

## Manuscript and figures

The current PDF is distributed alongside this source archive in the data repository under `release/RefSeg_CA_MVA.pdf`. Install the official Springer Nature LaTeX template and normal TeX Live packages, placing sn-jnl.cls and sn-mathphys-num.bst alongside the manuscript as instructed by Springer. The repository does not redistribute these third-party generated files.

```bash
python code/figures_r3.py
python code/build_latex.py
```

The synthetic download is required before regenerating image panels. All written figures are monochrome vector graphics; the experimental raster inputs retain their colors.

## Human evaluation

Read review/HUMAN_VALIDATION_PROTOCOL.md and review/INSAN_DEGERLENDIRMESI_UYGULAMA_KILAVUZU.md. Two independent readers, pre-adjudication agreement, adjudicated semantic labels, and a separate image-aware harm audit are still needed. Automated checks are not human labels.

## Provenance and reuse

See REPRODUCIBILITY.md, THIRD_PARTY.md, CITATION.cff, and SHA256SUMS. This public deposit establishes accessibility. Authors have not yet selected an additional blanket reuse license for original material; existing per-file third-party terms continue to apply.

Browse all source files at https://huggingface.co/datasets/Ethosoft/RefSeg-CA/tree/main/source. This organization-owned Hugging Face source tree is the code host for this release.