Skip to main content
Back to timeline
arXivSource publication:

Remote-sensing segmentation stops adopting every revision: GeoReVis keeps evidence from rejected masks and reaches 72.00% gIoU on LaSeRS

Synopsis

Language-guided remote-sensing segmentation often revises masks through feedback, yet a revision can damage regions that were already correct, and even the best whole mask can omit correct fragments held by weaker candidates. GeoReVis unifies execution-verified revision, candidate selection and cross-mask evidence consolidation: it adopts only corrections that improve the executed mask, partitions the spatial disagreements among candidates into composable regions, and jointly decides which regions to keep under the full instruction and image context. On LaSeRS it reaches 72.00% gIoU and 77.01% cIoU, and on ReVis-Eval it lowers the degradation rate from 17.90% for SocioReasoner∗ to 6.59%.

AI-generated editorial illustration: GeoReVis: Learning Selective Revision and Evidence Consolidation for Remote Sensing Segmentation

Interpretation

GeoReVis changes feedback revision in remote-sensing segmentation from adopt-on-proposal to compare-after-execution, and lets rejected candidate masks keep contributing local evidence. Earlier feedback-revision methods send a predicted mask back to the model for correction, or compare several predictions to pick a better whole mask; the analysis shows revision lowers IoU by more than 0.01 on 27.81% and 19.34% of LaSeRS samples, and even the highest-IoU candidate stays below a regional upper bound that draws on target evidence from all candidates. Across 1,158 LaSeRS samples, the optimal candidate reaches 66.94% gIoU for GeoReVis-R against a 75.39% regional oracle, and 63.90% against 68.97% for SocioReasoner∗; on ReVis-Eval, GeoReVis records a 5.28% IoU gain and a 6.59% degradation rate, versus 1.24% and 17.90% for SocioReasoner∗.

Each round's revision actions, candidate comparisons and region composition run through role-specific adapters on one 2B-scale vision-language model, while a frozen segmenter only turns prompts into masks. A revision is written as a replace, add or erase action with boxes and signed points, executed by the segmenter to produce a candidate; candidate selection compares full-mask and difference views under the complete instruction to choose a reference mask. Cross-mask evidence consolidation splits the deduplicated spatial disagreements among candidates into eight-connected regions by membership signature, jointly predicts whether each region is kept under the full instruction and image context, then compares the reconstruction with the reference mask before adoption; on the 3,474 candidate banks of ReVis-Eval, GeoReVis reaches 70.41% IoU against 69.37% for the optimal candidate.

Training signals come from what the model actually executes, not only from whole-mask IoU alignment. An execution-driven data pipeline labels which corrections are worth learning, which candidates are better and which regions should be combined, using the measured IoU change after real execution, extending supervision to revision actions, candidate preferences and region composition. In the LaSeRS component ablation, execution-verified revision raises initial segmentation from 63.28% to 64.64% gIoU while unverified revision lowers it to 62.71%; adding candidate selection cuts the degradation rate from 27.81% to 10.36%, and adding evidence consolidation brings gIoU to 72.00%.

Perspective

The demonstrated scope is language-guided remote-sensing segmentation on five benchmarks, LaSeRS, EarthReason, RRSIS-D, RefSegRS and RISBench, plus the ReVis-Eval process evaluation built from held-out LaSeRS scenes; one GeoReVis model is trained per benchmark, and inference uses three rounds by default.

Consolidation is restricted to the union of observed candidates, so a target region absent from initial segmentation and every revision round cannot be recovered by consolidation; gains saturate after three rounds, with a fourth raising gIoU only from 72.00% to 72.06% while per-query time grows from 3.98 to 5.99 seconds; some EarthReason annotations have coarse boundaries, which adds noise to execution-driven supervision.