GeoReVis: Learning Selective Revision and Evidence Consolidation for Remote Sensing Segmentation
Organizations: Beijing Institute of Technology
Abstract
Recent methods try to inspect and revise masks to mitigate referential ambiguity and localization errors caused by complex spatial relations in language-guided remote sensing segmentation. However, these solutions often lead to Revision Degradation: revising missing regions may introduce errors into previously correct ones. Even a globally better mask can omit correct regions found in inferior masks. We term this Evidence Fragmentation: correct target evidence is scattered across imperfect masks. All masks generated by the model at this time are suboptimal. We propose GeoReVis, which unifies revision guided by execution feedback, Candidate Selection, and Cross-Mask Evidence Consolidation. It learns fine-grained revision from real actions that improve segmentation when executed and learns to select a better mask through quality comparisons between actual candidates, reducing the risk of degradation during revision. The selected whole mask can still discard useful regions. GeoReVis therefore retains all candidates across rounds, partitions their spatial disagreements into composable regions, and jointly reconstructs regions conditioned on the full instruction and context. This makes complementary and fragmentary evidence reusable regardless of its source candidate's overall quality, yielding improved segmentation. We further build a data pipeline to generate fine-grained multi-round segmentation trajectories and derive supervision for revision actions, candidate selection, and region composition from the model's own outputs, extending training beyond whole-mask IoU supervision. We also propose ReVis-Eval, a benchmark protocol evaluating multi-round revision and cross-round evidence consolidation accuracy. Experiments on ReVis-Eval and 5 benchmarks for remote sensing reasoning and referring segmentation demonstrate the effectiveness of GeoReVis.
Figures & tables
| Method | Backbone | Segmenter | Granularity | Multiplicity | Reasoning | Length | Overall | ||||||
| Sem. | Inst. | Part | Single | Multi | Expl. | Impl. | Short | Long | gIoU | cIoU | |||
| Single-pass Segmentation Models | |||||||||||||
| LISA [ 15 ] | LLaVA-7B | SAM | 26.40 | 20.50 | 16.10 | 37.30 | 18.20 | 27.10 | 21.50 | 34.10 | 38.40 | 26.60 | 25.10 |
| M 2 A [ 12 ] | LLaVA-7B | SAM | 30.10 | 23.00 | 18.60 | 45.40 | 20.90 | 35.80 | 23.30 | 35.80 | 41.50 | 30.50 | 29.30 |
| GLaMM-ft [ 26 ] | Vicuna-7B | SAM | 44.80 | 41.20 | 32.60 | 47.30 | 32.20 | 59.10 | 42.60 | 50.40 | 42.60 | 43.60 | 48.30 |
| GeoPixel [ 31 ] | IXC-2.5-7B | SAM2 | 51.40 | 44.10 | 43.90 | 55.00 | 49.20 | 66.50 | 41.10 | 51.10 | 51.40 | 50.40 | 55.20 |
| Method | Val | Test | Avg. | ||
|---|---|---|---|---|---|
| gIoU | cIoU | gIoU | cIoU | ||
| LISA [ 15 ] | 61.04 | 57.39 | 60.88 | 59.10 | 59.60 |
| RemoteReasoner [ 36 ] | 69.02 | 67.80 | 70.96 | 69.13 | 69.23 |
| SegEarth-R2 [ 35 ] | 72.30 | 68.10 | 73.50 | 69.50 | 70.85 |
| UniGeoSeg [ 24 ] | 72.54 | 73.30 | 73.08 | 74.61 | 73.38 |
| Think2Seg-RS [ 40 ] | 72.16 | 72.87 | 73.36 | 75.60 | 73.50 |
| Method | RRSIS-D | RefSegRS | RISBench |
| Referring Segmentation Specialists | |||
| RS2-SAM 2 ‡ [ 29 ] | 66.72 | 73.90 | 68.55 |
| RSRefSeg 2 [ 4 ] | 69.17 | 77.39 | 72.57 |
| CROSS ‡ [ 23 ] | 68.92 | 79.51 | 70.86 |
| Generative VLM-Based Models | |||
| SocioReasoner ∗ [ 34 ] | 64.61 | 71.94 | 67.62 |
| Method | IoU | FG | DR | CP |
|---|---|---|---|---|
| SegAgent ∗ | 0.36 | +0.43 | 30.68 | 87.36 |
| SAM-Veteran ∗ | +0.86 | +0.71 | 19.43 | 92.18 |
| SocioReasoner ∗ | +1.24 | +0.86 | 17.90 | 93.04 |
| GeoReVis-R | +0.39 | +1.12 | 31.15 | 88.27 |
| GeoReVis-RS | +4.94 | +1.95 | 7.72 | 95.37 |
| GeoReVis | +5.28 | +2.08 | 6.59 | 96.12 |
| Method | IoU | BWG | BWG | HR | AP |
|---|---|---|---|---|---|
| Union | 60.26 | 9.11 | 3.37 | 100.00 | 35.61 |
| Majority Vote | 67.65 | 1.72 | 0.93 | 21.94 | 71.36 |
| Whole-Mask Selection | 68.74 | 0.63 | 1.46 | 19.47 | 68.15 |
| Independent Regions | 68.66 | 0.71 | +1.11 | 38.63 | 66.29 |
| Current Round | 69.60 | +0.23 | +1.72 | 41.35 | 76.82 |
| GeoReVis | 70.41 | +1.04 | +2.61 | 48.17 | 79.54 |
| Components | LaSeRS | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| U.R. | V.R. | C.S. | C.C. | M.F. | E.C. | gIoU | cIoU | DR | OR |
| 63.28 | 62.85 | – | – | ||||||
| 62.71 | 62.93 | 32.47 | 37.05 | ||||||
| 64.64 | 66.65 | 27.81 | 36.44 | ||||||
| 66.58 | 70.62 | 10.36 | 33.16 | ||||||
| 67.90 | 72.81 | 8.72 | 35.92 | ||||||
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
| Quantity | GeoReVis-R | SocioReasoner ∗ |
|---|---|---|
| LaSeRS macro gIoU: main Fig. 1 | ||
| Initial Segmentation | 63.28 | 62.26 |
| Revised | 64.64 | 63.13 |
| Optimal Candidate | 66.94 | 63.90 |
| Regional Oracle | 75.39 | 68.97 |
| With Full GeoReVis | 72.00 | 66.45 |
| Hyperparameter | Value | |
|---|---|---|
| Shared Settings | ||
| Backbone | Qwen3.5-2B | |
| Segmenter | SAM3 (frozen) | |
| Optimizer | AdamW | |
| Effective batch size | 16 | |
| Weight decay | 0 | |
| Method | Semantic | Instance | Part | Single | Multi | Explicit | Implicit | Short | Long | Mean |
|---|---|---|---|---|---|---|---|---|---|---|
| Single-pass Segmentation Models | ||||||||||
| LISA [ 15 ] | 26.40/23.20 | 20.50/25.00 | 16.10/11.60 | 37.30/32.20 | 18.20/22.40 | 27.10/24.30 | 21.50/25.60 | 34.10/27.80 | 38.40/33.90 | 26.60/25.10 |
| PixelLM [ 28 ] | 32.00/32.80 | 26.60/30.00 | 13.20/16.50 | 44.30/40.40 | 20.20/23.50 | 25.00/23.10 | 23.90/21.90 | 41.60/38.90 | 37.10/34.50 | 29.30/29.10 |
| M 2 A [ 12 ] | 30.10/33.00 | 23.00/24.80 | 18.60/17.20 | 45.40/37.60 | 20.90/24.80 | 35.80/30.40 | 23.30/26.70 | 35.80/32.80 | 41.50/36.70 | 30.50/29.30 |
| GLaMM-ft [ 26 ] | 44.80/47.90 | 41.20/48.30 | 32.60/42.70 | 47.30/50.30 | 32.20/41.00 | 59.10/60.30 | 42.60/44.80 | 50.40/54.80 | 42.60/44.80 | 43.60/48.30 |
| GeoPixel [ 31 ] | 51.40/57.20 | 44.10/49.30 | 43.90/52.40 | 55.00/45.80 | 49.20/49.70 | 66.50/61.30 | 41.10/58.30 | 51.10/59.30 | 51.40/63.20 | 50.40/55.20 |
| Method | Val | Test | ||
|---|---|---|---|---|
| gIoU | cIoU | gIoU | cIoU | |
| GeoPixel [ 31 ] | 52.13 | 54.23 | 52.53 | 53.90 |
| PixelLM [ 28 ] | 57.94 | 57.79 | 60.01 | 59.22 |
| LISA [ 15 ] | 61.04 | 57.39 | 60.88 | 59.10 |
| VisionReasoner [ 21 ] | 66.58 | 67.13 | 67.19 | 65.41 |
| SegEarth-R1 [ 16 ] | 68.60 | 64.13 | 70.75 | 68.25 |
| Method | RRSIS-D | RefSegRS | RISBench | ||
|---|---|---|---|---|---|
| gIoU | cIoU | gIoU | cIoU | gIoU | |
| LISA ‡ [ 15 ] | 26.78 | 27.94 | 24.62 | 35.17 | 28.36 |
| PixelLM ‡ [ 28 ] | 31.65 | 33.60 | 28.95 | 39.48 | 32.47 |
| RMSIN [ 19 ] | 64.20 | 77.79 | 62.58 | 75.72 | 63.07 |
| VisionReasoner ‡ [ 21 ] | 63.47 | 73.85 | 69.71 | 76.93 | 66.41 |
| RemoteReasoner ‡ [ 36 ] | 64.12 | 74.36 | 71.28 | 77.64 | 66.95 |
| Configuration | gIoU | cIoU |
|---|---|---|
| Initial Segmentation | 71.46 | 70.38 |
| Selective Revision (1 Round) | 74.21 | 76.02 |
| + Multi-Round Feedback | 75.37 | 77.24 |
| + Evidence Consolidation | 79.28 | 81.17 |
| Components | LaSeRS | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| U.R. | V.R. | C.S. | C.C. | M.F. | E.C. | gIoU | cIoU | DR | OR | Frag. |
| 63.28 | 62.85 | – | – | 51.36 | ||||||
| 62.71 | 62.93 | 32.47 | 37.05 | 50.28 | ||||||
| 64.64 | 66.65 | 27.81 | 36.44 | 51.02 | ||||||
| 66.58 | 70.62 | 10.36 | 33.16 | 54.47 | ||||||
| 67.90 | 72.81 | 8.72 | 35.92 | 54.72 | ||||||
| Setting | gIoU | cIoU |
| Full objective | 72.00 | 77.01 |
| w/o whole-mask loss | 71.71 | 76.58 |
| : target-average term only | 72.04 | 76.67 |
| : pooled term only | 71.86 | 77.12 |
| w/o region loss | 71.93 | 76.95 |
| w/o adoption loss | 71.82 | 76.87 |
| Integration Strategy | gIoU | cIoU | BWG P |
| Whole-Mask Selection | 69.19 | 73.43 | 1.46 |
| Majority Vote | 68.41 | 72.80 | 0.93 |
| Union | 66.84 | 71.93 | 3.37 |
| Subset Union | 68.99 | 73.19 | +0.28 |
| Independent Regions | 69.17 | 74.05 | +1.11 |
| Current Round Only | 71.44 | 76.32 | +1.72 |
| Rounds | 1 | 2 | 3 | 4 |
|---|---|---|---|---|
| GeoReVis-RS | 67.90/72.81 | 68.75/73.45 | 69.19/73.43 | 69.14/73.40 |
| GeoReVis | 68.13/72.84 | 70.26/74.95 | 72.00/77.01 | 72.06/76.95 |
| Time (s/query) | 2.12 | 3.98 | 5.99 | 7.91 |
| Method | Time (s/query) | Calls | Memory (GB) |
|---|---|---|---|
| SocioReasoner ‡ | 2.81 | 3.00 | 14.98 |
| SAM-Veteran ‡ | 5.41 | 4.07 | 24.68 |
| SegAgent ‡ | 6.68 | 7.64 | 23.23 |
| GeoReVis w/o consolidation | 5.25 | 9.08 | 19.35 |
| GeoReVis | 5.99 | 9.08 | 20.65 |
| Method | IoU | FG | DR | CP | NER |
|---|---|---|---|---|---|
| SegAgent ∗ | 0.36 | +0.43 | 30.68 | 87.36 | 1.62 |
| SAM-Veteran ∗ | +0.86 | +0.71 | 19.43 | 92.18 | 4.27 |
| SocioReasoner ∗ | +1.24 | +0.86 | 17.90 | 93.04 | 5.83 |
| GeoReVis-R | +0.39 | +1.12 | 31.15 | 88.27 | 2.84 |
| GeoReVis-RS | +4.94 | +1.95 | 7.72 | 95.37 | 17.86 |
| GeoReVis | +5.28 | +2.08 | 6.59 | 96.12 | 19.43 |
| Method | IoU | BWG | BWG | HR | AP |
|---|---|---|---|---|---|
| Keep Reference | 66.50 | 2.87 | 6.63 | 0.00 | – |
| SAM Score | 66.17 | 3.20 | 6.09 | 8.71 | 57.42 |
| Prompted Qwen | 64.69 | 4.68 | 6.37 | 15.86 | 49.08 |
| Union | 60.26 | 9.11 | 3.37 | 100.00 | 35.61 |
| Majority Vote | 67.65 | 1.72 | 0.93 | 21.94 | 71.36 |
| Whole-Mask Selection | 68.74 | 0.63 | 1.46 | 19.47 | 68.15 |
| Method | GeoReVis | VisionReasoner ‡ | SocioReasoner ‡ |
|---|---|---|---|
| SegAgent ∗ | 0.83 / 33.50 | +0.12 / 27.62 | 0.41 / 31.16 |
| SAM-Veteran ∗ | +0.41 / 21.84 | +1.47 / 16.88 | +0.66 / 19.78 |
| SocioReasoner ∗ | +0.74 / 19.29 | +1.86 / 15.88 | +1.09 / 18.66 |
| GeoReVis-R | 0.12 / 33.99 | +1.03 / 28.43 | +0.21 / 31.25 |
| GeoReVis-RS | +3.62 / 8.91 | +6.37 / 6.50 | +4.71 / 7.84 |
| GeoReVis | +3.88 / 7.64 | +6.81 / 5.32 | +5.02 / 6.90 |
| Method | BWG P | BWG | BWG | CP E |
|---|---|---|---|---|
| Keep Reference | 6.63 | 1.72 | 2.87 | 100.00 |
| SAM Score | 6.09 | 2.32 | 3.20 | 92.63 |
| Prompted Qwen | 6.37 | 4.16 | 4.68 | 89.27 |
| Union | 3.37 | 10.86 | 9.11 | 100.00 |
| Majority Vote | 0.93 | 1.96 | 1.72 | 97.62 |
| Whole-Mask Selection | 1.46 | 0.38 | 0.63 | 95.84 |
| Method | Backbone | Segmenter | LaSeRS | EarthReason | RRSIS-D | RefSegRS | RISBench | Avg. |
|---|---|---|---|---|---|---|---|---|
| SOTA Methods | ||||||||
| GeoReason [ 9 ] | Qwen3-VL-4B | SAM2 | 64.96 | 78.11 | 67.32 | 76.38 | 71.36 | 71.63 |
| UniGeoSeg [ 24 ] | Phi-1.5 + Swin-B | Mask2Former | 72.81 | 69.25 | 76.06 | 70.94 | ||
| CROSS [ 23 ] | SigLIP 2 | SAM2 | – | – | 68.92 | 79.51 | 70.86 | – |
| RSRefSeg 2 [ 4 ] | SigLIP 2 | SAM2 | – | – | 69.17 | 77.39 | 72.57 | – |
| GeoReVis | Qwen2.5-VL-7B | SAM2 | 70.94 | 78.20 | 69.51 | 79.82 | 72.63 | 74.22 |
Remote Sensing Image Segmentation
Remote sensing image segmentation labels land cover and objects in satellite, aerial, radar, and multispectral imagery.
See all 235 papers