GIC-SAM: Coupling Geometric Modeling with Instance Reasoning for Dense Instance Segmentation
Year: 2026
Type: conference-paper
Abstract: Instance segmentation in dense scenes is a challenging problem due to severe object crowding, ambiguous boundaries, and frequent overlaps, which are commonly encountered in applications such as histopathological analysis, remote sensing, and industrial inspection. Existing approaches, including prompt-free and prompt-guided methods, often suffer from excessive false positives and inaccurate boundaries due to conflicting modeling signals between pixel-level geometry and instance-level semantics: prompt-free methods emphasize local geometric consistency but lack instance discrimination, while prompt-guided approaches capture global instance priors yet overlook fine-grained boundary details. Consequently, these paradigms fail to achieve a balanced integration of geometric precision and instance-level reasoning. To address this issue, we propose GIC-SAM, a unified framework that explicitly couples geometric modeling with instance-level reasoning. Specifically, we propose a Semantic-aware Multi-granularity Prompter (SMP) to jointly encode geometric structures and semantic cues, generating boundary-aware representations and instance-aware prompts for accurate instance perception. Building upon this, a Multi-Level Mask Fusion Strategy (MLFS) is introduced to perform instance-level validation and cross-level refinement, enabling more accurate instance predictions by suppressing false positives and improving boundary quality. Additionally, we propose a Robust Prompt Training (RPT) strategy to mitigate the discrepancy between training and inference prompts by explicitly modeling prompt uncertainty, thereby enhancing robustness. Extensive experiments on four dense-scene benchmarks demonstrate that our method consistently outperforms state-of-the-art approaches, achieving superior boundary quality, fewer false positives, and improved robustness in complex scenarios.
Institution: Soochow University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7225314366