File size: 3,151 Bytes
8c2104c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
---
language:
- en
base_model: Qwen/Qwen3.5-2B
pipeline_tag: image-segmentation
tags:
- remote-sensing
- referring-segmentation
- reasoning-segmentation
- lora
- georevis
---

# GeoReVis

**GeoReVis: Learning Selective Revision and Evidence Consolidation for Remote Sensing Segmentation**

This repository contains the final GeoReVis weights trained on the full eligible LaSeRS training set. GeoReVis combines execution-feedback revision, Candidate Selection, and Cross-Mask Evidence Consolidation over three rounds.

The model uses Qwen3.5-2B with role-specific LoRA adapters and a frozen SAM3 segmentation decoder. Its final LaSeRS paper results are **72.00% gIoU** and **77.01% cIoU**.

## Files

| Directory | Role |
|---|---|
| `planner/` | Target Decomposition adapter |
| `initializer/` | Merged Spatial Grounding model and processor |
| `reviser/` | Execution-Verified Feedback Revision adapter |
| `selector/` | Candidate Selection adapter |
| `critic/` | Feedback Assessment adapter |
| `feedback_reviser/` | Multi-round Feedback Revision adapter |
| `transition_selector/` | Multi-round Candidate Selection adapter |
| `consolidation/adapter/` | Final Cross-Mask Evidence Consolidation adapter |
| `consolidation/readout.pt` | Joint regional readout, including its frozen initialization branch |
| `inference.json` | Portable configuration for the GeoReVis code release |
| `SHA256SUMS` | File checksums |

All listed roles are used in the final system. The consolidation adapter alone is not a standalone segmentation model. Training checkpoints, optimizer states, logs, and evaluation outputs are not included.

## Usage

Use these weights with the **GeoReVis_OpenSource** code release. Download this repository into `GeoReVis_FinalWeights/` beside the code directory:

```bash
hf download georevis/GeoReVis --local-dir GeoReVis_FinalWeights
```

```text
workspace/
  GeoReVis_OpenSource/
  GeoReVis_FinalWeights/
```

Follow the code README to install the environment and prepare the external [Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B) backbone, [SAM3](https://huggingface.co/facebook/sam3), and TRAIN visual reference library. These external assets are not included in this repository. The reference library is constructed by the Execution-Driven Data Pipeline.

Run the code's inference launcher with `--config ../GeoReVis_FinalWeights/inference.json`. The configuration resolves `${project}` to the code checkout and specifies all role weights explicitly.

```bash
cd GeoReVis_OpenSource
python scripts/run_inference.py \
  --config ../GeoReVis_FinalWeights/inference.json --gpus 0,1,2,3 \
  --inputs datasets/manifests/lasers/test/inputs.jsonl \
  --output outputs/pretrained/lasers_test
```

The Execution-Driven Data Pipeline uses **MiniMax-M3.1** to improve training-data quality. It is not needed at inference once the trained weights and reference library are available.

## Integrity and Base Model Terms

To verify downloaded files, run `sha256sum -c SHA256SUMS` in the weights directory. The Qwen3.5-2B base-model license is retained in `LICENSE-Qwen3.5.txt`. SAM3 and datasets retain their own distribution terms.