Training-Free · Remote Sensing · Referring Segmentation

GeoSelect

Spatial-Program Execution for Training-Free Referring Remote Sensing Image Segmentation

Yuhang Jiang1, Guohui Deng1,2,*, Miaozhong Xu3, Chao Ruan1, Jinling Zhao1, Linsheng Huang1

1School of Internet, Anhui University
2National Engineering Research Center for Agro-Ecological Big Data Analysis & Application, Anhui University
3State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University

*Corresponding author

Under Review
58.86
RRSIS-D mIoU
best training-free result
×2.08
over the best prior
training-free method
55.27
RISBench mIoU
frozen config, transferred
×1.74
over the best prior
training-free method
0
trained parameters
every component frozen

Abstract

Referring remote sensing image segmentation (RRSIS) segments the single object that an aerial image's natural-language expression refers to. Existing training-free methods resolve the expression through implicit vision–language activations or region–text similarity, which gives weak control over the spatial, comparative, and ordinal relations that dominate aerial referring and cannot represent constructions such as the largest ship or the second court from the left. We propose GeoSelect, a training-free pipeline that reframes referring as the execution of a typed spatial program. A frozen, text-only language model synthesises a program in a small domain-specific language from the expression alone; a well-formedness checker accepts it; and a deterministic executor evaluates it over a single scored candidate set type that unifies continuous geometric fields with discrete set and order operators (extremum, ordinal, counted union, binary relation). A reliability ladder degrades any failing program to a field-only selector, so a malformed program falls back rather than forfeiting an answer. GeoSelect attains 58.86 mIoU / 59.48 oIoU on RRSIS-D test (more than twice the best prior training-free method and about 93% of the supervised specialist RMSIN with no training), and the identical configuration transfers to RISBench at 55.27 mIoU. Holding candidates and the segmenter fixed, explicit program execution substantially improves the selector over implicit matching, attributing the gain to the executed representation rather than the backbone. An image-level exposure audit confirms the result is not inflated by detector–benchmark overlap.

Why Explicit Execution?

Same candidate boxes, same segmenter; only the selector differs. A real RRSIS-D example: “the airplane on the right”, with several same-class candidates.

selector result

GeoSelect executes the program and picks the airplane matching the ground truth ✓

scored candidate set

The executed program scores every candidate; brighter is a higher score, and the chosen box (green) wins.

Predicted mask Ground truth Candidate boxes Chosen box

Method

A typed spatial program, synthesised then executed over one scored candidate set; every component is frozen and off-the-shelf.

C1

Explicit Spatial-Program Execution

A frozen text-only LLM synthesises a typed program in a small DSL from the expression alone; a well-formedness checker accepts it; a deterministic executor runs it bottom-up. Every intermediate program, field, and ranking is inspectable.

C2

Fields + Discrete Operators

One scored-candidate-set type unifies continuous geometric fields (position, proximity) with discrete set and order operators (extremum, ordinal, counted union, relation) that express the superlative, ordinal, and compositional constructions a closed lexicon cannot.

C3

Reliability Ladder

An ill-formed or failing program falls back to the field-only selector on the same candidates, which is recovered exactly; in aggregate the full system stays at or above this field-only floor, so a malformed program never forfeits an answer.

Trace the Pipeline

A real, frozen-config trace; click each stage. The program is executed over one scored candidate set, so every intermediate is inspectable.

Main Results

Supervised specialists and trained VLMs are an upper reference, not training-free competitors.

RRSIS-D test (n = 3,481)

MethodVenuemIoUoIoU
Fully-supervised specialists
LAVTCVPR'2261.1276.48
LGCETGRS'2460.9876.33
RMSINCVPR'2463.3876.55
FIANetTGRS'2464.0176.81
CroBIMarXiv'2564.2476.37
ProVGarXiv'2665.4477.62
BTDNetarXiv'2566.0479.23
RSRefSeg2TGRS'2669.1779.45
Trained vision–language models
GeoGroundarXiv'2460.50
Text4Seg++TPAMI'2662.8074.20
SegEarth-R1arXiv'2566.4078.01
Sosa et al. (LoRA)arXiv'2667.6
Training-free
SAM3arXiv'2615.5016.87
EKP-HRMGRSL'2619.2120.68
DGL-RSISJAG'2621.50
Sosa et al. (zero-shot)arXiv'2624.9
RSVG-ZeroOVAAAI'2628.3522.83
GeoSelect (ours)58.8659.48

×2.08 over the best prior training-free method (RSVG-ZeroOV) and ×3 over EKP-HRM, about 93% of the supervised specialist RMSIN, with no training.

RISBench test (n = 16,159, frozen RRSIS-D config)

MethodVenuemIoUoIoU
Fully-supervised
RRNCVPR'1843.1849.67
BRINetCVPR'2042.9148.73
CMPC+TPAMI'2146.7353.98
CRISCVPR'2255.1869.11
LAVTCVPR'2261.9374.15
RMSINCVPR'2463.0774.09
CARISMM'2365.7975.10
CroBIMarXiv'2567.3273.61
RSRefSeg2TGRS'2672.5774.77
Training-free
SAM3arXiv'2621.0926.20
RSVG-ZeroOVAAAI'2631.8426.35
GeoSelect (ours)55.2755.88

A single frozen configuration, selected before test evaluation, transfers directly, at ×1.74 the best prior training-free method, surpassing the early LSTM-based supervised methods RRN, BRINet, and CMPC+ with no test-set recalibration.

Key Findings

C1

Referring as explicit program execution

A typed program, synthesised by a frozen text-only LLM and executed over one scored-candidate-set type that unifies continuous fields with discrete operators, reaches 58.86 mIoU on RRSIS-D, ×2.08 the prior training-free baseline.

C2

The gain is representation, not backbone

With detection and segmentation held fixed, explicit program execution beats the stronger of two implicit selectors by +12.06 mIoU (+18.8 on image-relative, +0.9 on pure-attribute), a per-subset pattern that rules out detector or segmenter explanations.

C3

Transferable, and audited for leakage

The identical frozen configuration transfers across domains to RISBench, and an image-id-level seen/unseen exposure audit shows the leakage-robust unseen subset still reaches 58.4 mIoU, well above the best prior training-free method, so the detector overlap is a measured strength, not an unexamined confound.

Case Explorer

Real frozen-config outputs across the expression strata. Click any case for its synthesised program, per-stage trace, and ground truth.

BibTeX

@article{jiang2026geoselect,
  title   = {GeoSelect: Spatial-Program Execution for
             Training-Free Referring Remote Sensing Image Segmentation},
  author  = {Jiang, Yuhang and Deng, Guohui and Xu, Miaozhong and Ruan, Chao and Zhao, Jinling and Huang, Linsheng},
  journal = {arXiv preprint arXiv:2607.03869},
  year    = {2026}
}