Physics Guided Contrastive Pretraining for Material Attribute Learning from Impact Video Dynamics and Surface Deformations
INTERNATIONAL JOURNAL OF RESEARCH & SOCIAL SCIENCES · 10.06.2026 · 6 (1) · 213-225
DOI: 10.66238/ijrss.v6i1.109 ↗Kurz gefasst
Fachartikel in der Zeitschrift INTERNATIONAL JOURNAL OF RESEARCH & SOCIAL SCIENCES, online erschienen am 10.06.2026. Verfasst von Logan Douglas und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
The ability to infer the physical and material properties of objects from visual observations is a fundamental capability for autonomous agents, robotics, and interactive systems. Traditional visual material recognition systems predominantly rely on static surface features such as color, texture, and reflectance, which often fail to capture the intrinsic physical attributes like elasticity, density, and friction. Impact videos, which capture the dynamic interaction between objects, provide rich visual cues regarding these hidden material properties through structural deformation, rebound trajectories, and vibrational responses. In this paper, we propose a novel framework for material attribute learning from impact videos employing physics-guided contrastive pretraining. Unlike standard contrastive learning approaches that rely on arbitrary data augmentations, our methodology integrates physical principles to structure the representation space. We introduce a physics-aware sampling strategy that pairs video sequences based on underlying physical similarities rather than mere visual resemblance. By optimizing a contrastive objective modulated by physical divergence, the model learns a representation space where distances reflect material property variations. Through extensive experiments on both synthetic and real-world impact datasets, the proposed architecture demonstrates significant improvements over baseline methods in predicting coefficients of restitution, surface friction, and mass distributions. The findings suggest that integrating domain-specific physical constraints into self-supervised learning paradigms bridges the gap between superficial visual features and intrinsic mechanical properties, paving the way for more robust and generalizable physical scene understanding.
Zitieren
APA
Douglas, L., Matsuda, R., & Kuroda, A. (2026). Physics Guided Contrastive Pretraining for Material Attribute Learning from Impact Video Dynamics and Surface Deformations. INTERNATIONAL JOURNAL OF RESEARCH & SOCIAL SCIENCES, 6(1), 213-225. https://doi.org/10.66238/ijrss.v6i1.109
Vancouver
Douglas L, Matsuda R, Kuroda A. Physics Guided Contrastive Pretraining for Material Attribute Learning from Impact Video Dynamics and Surface Deformations. INTERNATIONAL JOURNAL OF RESEARCH & SOCIAL SCIENCES. 2026;6(1):213-225. doi:10.66238/ijrss.v6i1.109
BibTeX
@article{Douglas2026,
author = {Douglas, Logan and Matsuda, Rina and Kuroda, Asuka},
title = {Physics Guided Contrastive Pretraining for Material Attribute Learning from Impact Video Dynamics and Surface Deformations},
journal = {INTERNATIONAL JOURNAL OF RESEARCH & SOCIAL SCIENCES},
year = {2026},
volume = {6},
number = {1},
pages = {213--225},
doi = {10.66238/ijrss.v6i1.109}
}
EUVIMED