Specialized structural pruning of foundation models for indoor robotic vision
Scientific Bulletin of UNFU · 26.02.2026 · 36 (1)
DOI: 10.36930/40360117 ↗Kurz gefasst
Fachartikel in der Zeitschrift Scientific Bulletin of UNFU, online erschienen am 26.02.2026. Verfasst von Б. П. Борківський und 1 weiteren Autor:innen. Die Arbeit nennt 21 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
We address the problem of deploying large-scale vision foundation models in real-time robotic systems, which is currently hindered by immense computational demands and significant inference latency. We analyze the limitations of general-purpose segmentation model compression methods, which frequently fail to maintain semantic precision and mask granularity when transferred to specialized environments such as indoor robotic navigation. We introduce a domain-specific structural optimization framework for segmentation models, designed to transform a heavyweight image segmentation architecture into a lightweight, high-precision tool tailored for specialized perception tasks. We examine the role of a multi-stage pipeline, beginning with a critical adaptation phase where a full-scale teacher model is specialized to establish robust prior knowledge of target textures and indoor geometries, providing a domain-aware baseline for knowledge transfer. We implement an alternate slimming framework that decomposes the vision transformer encoder into decoupled embedding and bottleneck internal dimensions, allowing sequential structural pruning that maintains architectural consistency for intermediate feature alignment. We integrate a robust data augmentation pipeline during the distillation and recovery phases, incorporating complex geometric and photometric transformations to stabilize the learning process and mitigate overfitting in data-scarce scenarios. Our results show that this comprehensive approach achieves a 73.3 % reduction in total trainable parameters and a 74.3 % reduction in Multiply-Accumulate operations (MACs), which effectively doubles the inference speed from 7 FPS to 15 FPS. We observe that the optimized student model experiences a marginal accuracy drop of only approximately 1 % compared to the specialized teacher model, consistently outperforming standard SlimSAM baselines in both semantic mask granularity and overall environmental robustness. We establish that our model demonstrates superior resistance to lighting variations and shadow interference, successfully addressing a common failure point in baseline architectures like FastSAM, that often distort segmentation masks under high-contrast indoor illumination. We conclude that the integration of domain-specific pruning and augmented distillation provides a superior and more stable path for enabling real-time, sophisticated computer vision on resource-constrained autonomous platforms.
Zitieren
APA
Борківський, Б. П., & Теслюк, В. М. (2026). Specialized structural pruning of foundation models for indoor robotic vision. Scientific Bulletin of UNFU, 36(1). https://doi.org/10.36930/40360117
Vancouver
Борківський БП, Теслюк ВМ. Specialized structural pruning of foundation models for indoor robotic vision. Scientific Bulletin of UNFU. 2026;36(1). doi:10.36930/40360117
BibTeX
@article{2026,
author = {Борківський, Б. П. and Теслюк, В. М.},
title = {Specialized structural pruning of foundation models for indoor robotic vision},
journal = {Scientific Bulletin of UNFU},
year = {2026},
volume = {36},
number = {1},
doi = {10.36930/40360117}
}
EUVIMED