Research map / Detection and segmentation / Open-set and few-shot detection
Rotated Multi-Scale Interaction Network for Referring Remote Sensing Image Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RMSIN, a rotation- and multi-scale-aware vision-language encoder-decoder for referring remote sensing image segmentation, and releases the RRSIS-D dataset of 17,402 image-caption-mask triplets.
Problem
Referring image segmentation in aerial imagery is hard due to extreme variation in object scales and orientations, which standard RIS methods handle poorly; existing RRSIS datasets are also small.
Approach
Combines intra-scale vision-language interaction (IIM) and cross-scale attention (CIM) in a compounded encoder, plus an Adaptive Rotated Convolution in the decoder that predicts multiple kernel rotation angles and weights to handle object orientation.
Key results
- RMSIN reaches 64.20 mIoU on the RRSIS-D test split, versus 61.04 for LAVT (+3.16).
- On the RRSIS-D validation split, RMSIN improves mIoU to 65.10 from 61.46 for LAVT (+3.64).
- Ablation: adding IIM and CIM together raises validation mIoU from 61.46 to 64.91 over the LAVT-style baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace LAVT with RMSIN | RRSIS-D · RRSIS-D val | mIoU: 61.46 -> 65.10 | Helped |
| replace LAVT with RMSIN | RRSIS-D · RRSIS-D test | mIoU: 61.04 -> 64.20 | Helped |
| add IIM while omitting CIM (ablation) | RMSIN · RRSIS-D val | mIoU: 61.46 -> 62.27 | Helped |
| add CIM while omitting IIM (ablation) | RMSIN · RRSIS-D val | mIoU: 61.46 -> 64.25 | Helped |
| add IIM and CIM together (ablation) | RMSIN · RRSIS-D val | mIoU: 61.46 -> 64.91 | Helped |
| add multi-scale attention to CIM (ablation) | RMSIN · RRSIS-D val | mIoU: 61.46 -> 61.46 | No clear effect |
| add feed-forward layer to CIM (ablation) | RMSIN · RRSIS-D val | mIoU: 61.46 -> 61.46 | No clear effect |
| add LRC to CIM (ablation) | RMSIN · RRSIS-D val | mIoU: 61.46 -> 64.25 | Helped |
| sum cross-stage features instead of concatenating them (ablation) | RMSIN decoder · RRSIS-D val | mIoU: 64.15 -> 62.76 | Hurt |
| replace ARC with static convolutions (ablation) | RMSIN decoder · RRSIS-D val | mIoU: 64.15 -> 61.56 | Hurt |
| replace one decoder convolution layer with ARC (ablation) | RMSIN decoder · RRSIS-D val | mIoU: 63.81 -> 62.53 | Hurt |
| replace two decoder convolution layers with ARC (ablation) | RMSIN decoder · RRSIS-D val | mIoU: 63.81 -> 62.48 | Hurt |
Limitations
- Evaluated only on the proposed RRSIS-D dataset; no other RRSIS benchmark reported.
- Annotation relies on SAM masks from RSVGD boxes, with manual refinement needed for problematic aerial images.
- Gains on the hardest P@0.9 metric are marginal relative to LAVT.
Experimental setup
Datasets: RRSIS-D RSVGD bounding boxes
Benchmarks: RRSIS-D
Metrics: P@0.5 P@0.6 P@0.7 P@0.8 P@0.9 oIoU mIoU
Baselines: RRN CSMA LSCM CMPC BRINet CMPC+ LGCE LAVT
Builds on: LAVT Swin Transformer BERT Segment Anything Model
Closest papers by meaning
- RS2-SAM2: Customized SAM2 for Referring Remote Sensing Image Segmentation (AAAI 2026 · similarity 0.86)
- Exploring Efficient Open-Vocabulary Segmentation in the Remote Sensing (AAAI 2026 · similarity 0.84)
- ROS-SAM: High-Quality Interactive Segmentation for Remote Sensing Moving Object (CVPR 2025 · similarity 0.83)
- ReSAM: Refine, Requery, and Reinforce: Self-Prompting Point-Supervised Segmentation for Remote Sensing Images (CVPR 2026 · similarity 0.83)
- Towards Open-Vocabulary Remote Sensing Image Semantic Segmentation (AAAI 2025 · similarity 0.81)
- GeoPixel: Pixel Grounding Large Multimodal Model in Remote Sensing (ICML 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.