Search papers, labs, and topics across Lattice.
This paper addresses the challenge of aircraft instance segmentation in SAR imagery by proposing a two-stage framework that combines SAR-to-optical image translation with an adapter-tuned Segment Anything Model (SAM). A diffusion model translates SAR aircraft slices into optical images to improve structural visibility, and then a SAM-adapter module generates instance and component-level masks on the translated images. A scattering-aware refinement module further refines the masks using SAR-specific scattering distributions, leading to improved fine-grained segmentation and strong zero-shot generalization.
Achieve fine-grained aircraft segmentation in SAR imagery by translating to the optical domain and adapting SAM, unlocking scalable automated mask annotation.
Instance segmentation in synthetic aperture radar (SAR) imagery has demonstrated notable success for certain target types such as vehicles and ships. However, accurate segmentation of aircraft in SAR images remains a significant challenge due to complex structural geometries, low-intensity and sparse backscattering, and frequent occurrences of incomplete or ambiguous contours. These inherent limitations hinder the generation of high-quality annotations and restrict downstream applications to coarse object detection tasks. To address these issues, this work proposes a two-stage framework combining SAR-to-optical image translation with an adapter-tuned segment anything model (SAM). In the first stage, a diffusion-based generative model first translates SAR aircraft slices into high-fidelity optical counterparts, enhancing structural visibility with continuous and interpretable contours. In the second stage, the SAM-adapter-based module produces instance-level and component-level masks on the translated images. In addition, to further improve alignment with SAR-specific characteristics, a scattering-aware refinement module refines the masks using the physical scattering distributions of the original SAR images. Experimental results demonstrate the effectiveness of the proposed framework in fine-grained segmentation and validate its strong zero-shot generalization ability, indicating its potential for scalable, automated mask annotation of aircraft in SAR imagery.
