Learning Where to Look: VLM-Assisted Efficient Inference for Sparse Aerial Objects Segmentation

作者：Fan Liu, Ting Wu, W Z Wang, Yifan Shi, Jiale Zhu, Liang Yao, Chengpeng Zhang · 发表于：IEEE Geoscience and Remote Sensing Letters · 年份：2026 · DOI：10.1109/lgrs.2026.3669375 · 研究领域：Advanced Neural Network Applications、Robotics and Sensor-Based Localization、Domain Adaptation and Few-Shot Learning

Existing high-performance semantic segmentation methods for remote sensing often struggle with the trade-off between inference efficiency and high-resolution processing, especially in scenarios characterized by sparse object distributions. In this paper, we present a vision-language model empowered framework tailored for efficient semantic segmentation in remote sensing. Technically, we introduce a Global Object Screening mechanism designed to prune redundant computations. By leveraging the image-text alignment capabilities of vision-language models, this module adaptively filters out background-dominated patches via a confidence-based greedy search strategy. Furthermore, to enhance local discriminability, we propose an efficient Semantic Feature Alignment method, utilizing contrastive learning on a constructed multi-scale image-text dataset to bolster fine-grained feature representation. In this manner, the semantic segmentation models can strike a good balance between identification accuracy and inference efficiency. Comprehensive evaluations on three standard benchmarks (iSAID, LoveDA, and Potsdam) reveal that our method delivers a dramatic increase in inference speed (up to 4.39× FPS) while still maintaining comparable segmentation accuracy. Our code and datasets are available at https://github.com/Ccwuting/VLM-Seg-Acceleration.