RefSeg-CA: command compliance and conservative repair

Research artifacts for RefSeg-CA: evaluating and repairing command compliance in generalized referring segmentation. This is an unpublished R3 manuscript release; independent human parser validation has not yet been conducted.

Contents

Artifact	Size / meaning
Procedural scenes	1,200 unique scene geometries, five seeds
Rendered images	2,400: matched flat and rich versions
Commands	28,800 command–image records; 12 per rendered image
Exact visible-instance maps	Shared geometry across the two render styles
Natural benchmark scores	All 49,492 official evaluation expressions, three model interfaces
Reviewer diagnostics	Four nested scope guards and 11 operating points for each frozen interface

The Dataset Viewer has 2,400 image rows, not 28,800 rows. commands_json contains the 12 commands, family, endpoint, and target-instance IDs for each image. instance_map is an indexed label image: 0 is background; the other integer values identify visible objects. A command mask is isin(instance_map, target_instance_ids); an empty ID list gives an empty mask.

Download exact experimental inputs

archives/rich-seed-*.tar.gz reconstruct the original data/rich/<scene_id>/ layout: flat/rich images, instance labels, and 12 command masks per scene. manifests/ contains complete synthetic records and a natural benchmark index. DATA_MANIFEST.json records SHA-256 values and byte counts.

results/primary-and-review-metrics.tar.gz contains the exact primary score records and reviewer replay traces. These reproduce the reported metrics without cluster access. They do not replace a pixel-level model rerun.

Browse the source code and editable manuscript file by file, or download RefSeg_CA_R3_Public_Source.zip. The current draft is RefSeg_CA_MVA.pdf. Code and data are hosted together under the Ethosoft organization on Hugging Face.

Construction and evaluation

Inputs are deterministically rendered with Pillow from scene graphs using seeds 11, 23, 37, 53, and 71. Images are 384×384 RGB. Shapes include circles, squares, triangles, stars, crosses, ellipses, capsules, hearts, diamonds, rings, and hexagons. Rich rendering changes texture, shading, and shadows while preserving target geometry. These are procedural raster inputs, not AI-generated illustrations or TikZ scenes.

Command families: attribute, relation, multiplicity, negative action, valid empty target, and paraphrase. The main paired-correctness table uses the first three and valid empty targets; action and paraphrase are analyzed separately. This is an evaluation-only split. Do not tune a proposed method on these records and then describe the same records as an untouched test set.

On the natural benchmark the primary scope guard accepts 3.15% of expressions. SFAP's pooled gIoU gains are approximately 0.108/0.072 percentage points for CLIPSeg/Grounded SAM. Harmful interventions remain non-negligible; a positive mean is not a per-example reliability guarantee. Trained ReLA is contextual, not a ReLA+SFAP experiment.

Natural data, models, and reuse

Original COCO image pixels and natural gRefCOCO expression text are not redistributed here. Obtain them from the gRefCOCO authors using revision 81eede59b3ac070049f597d023c0ff08d1fb80e9 and their documented image sources. The natural index preserves expression/reference/image IDs for reconstruction. Model weights also remain with their original publishers; preparation scripts pin revisions and verify hashes.

The authors have not yet assigned a blanket reuse license to original project code or procedural data. This public release provides access for inspection and reproducibility; it does not relicense upstream material. No DOI has been assigned.

Human validation status

No independent human annotations, inter-annotator agreement, or adjudicated parser-accuracy estimates are claimed. A 600-expression stratified protocol and a separate 120-case harmful-intervention audit are prepared. Reader expertise and dates must be actual observations, and AI-generated labels must not be described as human labels.

Citation

See CITATION.cff in the source archive. Cite the manuscript as an unpublished R3 research release and cite the GRES/gRefCOCO, CLIPSeg, Grounded SAM, and ReLA sources as appropriate for the components used.