Reproduce “Prompt-Calibrated SAM 3 for Open-Vocabulary Remote Sensing Semantic Segmentation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yanghui Song, Nanqing Liu, Haonan Yin, Yingjie Gao, Chengfu Yang, Qi Ming
ProC-SAM3 introduces a training-free and ground-truth-free framework that enhances SAM 3's open-vocabulary semantic segmentation for remote sensing imagery, achieving an average mIoU of 56.1% across eight benchmarks and outperforming prior methods by 3.9 percentage points.
Key insights
- Semantic alignment via robust prompt calibration, even without ground-truth masks, is critical for effectively adapting general-purpose foundation models like SAM 3 to the specific complexities of remote sensing imagery.
- Combining MLLM capabilities for initial prompt expansion with targeted category-specific constraints and matching effectively mitigates semantic drift while retaining beneficial prompt diversity.
- A carefully designed fusion strategy that selectively incorporates instance-level information, guided by confidence and agreement metrics, is essential for generating cleaner segmentations and preserving small object activations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.