Research map / Detection and segmentation / Open-set and few-shot detection
ROS-SAM: High-Quality Interactive Segmentation for Remote Sensing Moving Object
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.
Fine-tunes SAM with LoRA, unfrozen last encoder block, alternately-updated HQ-SAM decoders, and a crop/upsample inference pipeline to segment remote sensing moving objects in video satellite data.
Problem
Interactive (box-prompted) segmentation of small, ambiguous remote sensing video moving objects is poor with SAM, which suffers from domain gap, rough boundaries, and downsampling of small objects to 1024x1024.
Approach
Adapt SAM for remote sensing by combining LoRA on encoder Q/V with an unfrozen last encoder block, alternately updating SAM and HQ-SAM mask decoders, and redesigning training (large-scale jittering, rotation) and inference (prompt-centered cropping, 2x bicubic upsampling, single-object inference).
Key results
- On SAT-MTB, ROS-SAM reaches 50.54 IoU / 50.36 BIoU versus 43.41 for SAM with the inference pipeline and 37.25 for original SAM.
- The inference pipeline alone raises SAM IoU from 37.25 to 43.41 (over 6 points).
- On iSAID, ROS-SAM achieves 73.22 IoU versus 63.96 for HQ-SAM; on NPWS VHR-10 it achieves 87.46 IoU versus 78.44 for HQ-SAM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the proposed inference pipeline to SAM (ablation) | SAM · SAT-MTB | IoU: 37.25 -> 43.41 | Helped |
| Add the proposed inference pipeline to SAM2 (ablation) | SAM2 · SAT-MTB | IoU: 36.80 -> 41.75 | Helped |
| Use bicubic rather than bilinear interpolation at sampling rate 2 (ablation) | ROS-SAM inference pipeline · SAT-MTB | IoU: 41.58 -> 42.17 | Helped |
| Increase sampling rate from 2 to 4 with bicubic interpolation and one object at a time (ablation) | ROS-SAM inference pipeline · SAT-MTB | IoU: 43.41 -> 41.73 | Hurt |
| Add random rotation (ablation) | ROS-SAM training pipeline · SAT-MTB | IoU: 49.19 -> 49.43 | Helped |
| Add large-scale jittering (LSJ) (ablation) | ROS-SAM training pipeline · SAT-MTB | IoU: 49.43 -> 50.03 | Helped |
| Add both LSJ and random rotation (ablation) | ROS-SAM training pipeline · SAT-MTB | IoU: 49.19 -> 50.54 | Helped |
| Update the original SAM mask decoder (ablation) | ROS-SAM · SAT-MTB | IoU: 43.41 -> 42.82 | Hurt |
| Use the HQ mask decoder (ablation) | ROS-SAM · SAT-MTB | IoU: 43.41 -> 47.15 | Helped |
| Alternately update the HQ mask decoder and original SAM mask decoder (ablation) | ROS-SAM · SAT-MTB | IoU: 47.15 -> 48.16 | Helped |
| Fine-tune the image encoder using LoRA and/or the last encoder block (ablation) | SAM and ROS-SAM · SAT-MTB | IoU | Helped |
| Use ROS-SAM instead of HQ-SAM | SAM · iSAID | IoU: 63.96 -> 73.22 | Helped |
Surprising findings
- Directly updating the original SAM mask decoder decreased accuracy (42.82 vs 43.41 IoU), while updating only the HQ decoder helped more.
- Several other SOTA SAM fine-tuning methods (adapters and others) produced results below original SAM and were dropped from comparison.
Limitations
- Struggles to distinguish the number of airplane engines and objects visually ambiguous.
- Zero-shot evaluation on tracking datasets is qualitative only because no segmentation masks exist.
- Training uses only a subsampled SAT-MTB with some labels excluded; cars excluded from mask generation.
Experimental setup
Datasets: SAT-MTB iSAID NPWS VHR-10 SatSOT VISO OOTB
Benchmarks: SAT-MTB iSAID NPWS VHR-10 SatSOT VISO OOTB
Metrics: IoU BIoU
Baselines: SAM SAM2 HQ-SAM
Builds on: Segment Anything Model LoRA HQ-SAM
Closest papers by meaning
- SAMRS: Scaling-up Remote Sensing Segmentation Dataset with Segment Anything Model (NeurIPS 2023 · similarity 0.91)
- Exploring Efficient Open-Vocabulary Segmentation in the Remote Sensing (AAAI 2026 · similarity 0.89)
- Towards Open-Vocabulary Remote Sensing Image Semantic Segmentation (AAAI 2025 · similarity 0.88)
- ZoRI: Towards Discriminative Zero-Shot Remote Sensing Instance Segmentation (AAAI 2025 · similarity 0.87)
- InstructSAM: A Training-free Framework for Instruction-Oriented Remote Sensing Object Recognition (NeurIPS 2025 · similarity 0.86)
- SAM-I2V: Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost (CVPR 2025 · similarity 0.86)
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.