Teaching Feature-Based 2D-to-3D Reconstruction for Metaverse-Oriented Content Generation: An OpenCV Structure-from-Motion Methodology for Computer Science Students

Bahatskyi, Oleksii та Malakhov, Kyrylo та Bondarchuk, Andrii та Mashkina, Iryna та Nosenko, Tetiana (2026) Teaching Feature-Based 2D-to-3D Reconstruction for Metaverse-Oriented Content Generation: An OpenCV Structure-from-Motion Methodology for Computer Science Students Metaverse, 4 (7). ISSN 2810-9791

Текст O_Bahatskyi_K_Malakhov_A_Bondarchuk_I_Mashkina_T_Nosenko_Metaverse_7_4_KUBG.pdf Download (1MB)

Анотація

Feature-based Structure from Motion (SfM) is frequently taught through isolated OpenCV calls, which can hide the relationships among acquisition geometry, correspondence quality, pose recovery, reconstruction error, and downstream asset readiness. This article presents a constructively aligned methodology for teaching those relationships to undergraduate computer science students. The instructional artifact combines an instrumented OpenCV pipeline, explicit mathematical models, quality gates, a five-block laboratory sequence, an assessment rubric, reproducibility packaging, and a metaverse-content handoff. Technical characterization combines a five-view classroom timing baseline with three controlled rendered reference scenes—a textured object, a repeated-pattern facade, and a curved-form artifact approximated by ten planar facets—processed at 12, 24, and 48 views. The evaluation reports the number of cameras added to the sequential pose chain, mean adjacent-pair pose-validated inlier ratio (RANSAC and cheirality), reprojection root-mean-square error (RMSE), filtered point count, pose error against known trajectories, surface consistency against known geometry, and stage timings. Every adjacent camera pair yielded a pose, so all cameras were added to the sequential chain. The 24-view condition produced the largest point clouds (13,133–16,154 points), whereas 48 closely spaced views reduced the mean adjacent-pair pose-validated inlier ratio to 6.7%–22.1%, demonstrating that image count cannot compensate for insufficient parallax. Reprojection RMSE remained 0.34–0.44 pixels, which also shows why low residuals must be interpreted together with correspondence and pose diagnostics. Ground-truth checks exposed the same degeneracy: median translation-direction error reached 78.3°–179.3° in the 48-view conditions, and only 0.0%–22.0% of points were surface-consistent at δ = 0.12. The five-view timing baseline averaged 184.3 ms; feature processing and matching accounted for 46.4%. The sparse output is positioned as an inspectable upstream scaffold for dense reconstruction, meshing, texturing, semantic annotation, provenance, glTF packaging, and immersive-engine import. Because no student cohort was studied, the methodology specifies intended learning outcomes and a prospective evaluation protocol rather than claiming measured educational effectiveness.

Тип елементу :	Стаття
Ключові слова:	Metaverse; 3D Reconstruction; Computer Science Education; Structure from Motion; OpenCV; Feature Matching; Reproducible Computing; Semantic 3D Assets
Типологія:	Статті у періодичних виданнях > Інші (не входять ні до фахових, ні до баз даних)
Підрозділи:	Факультет інформаційних технологій та математики > Кафедра комп'ютерних наук
Користувач, що депонує:	доцент Тетяна Іванівна Носенко
Дата внесення:	26 Жов 2026 12:51
Останні зміни:	26 Жов 2026 12:51
URI:	https://elibrary.kubg.edu.ua/id/eprint/59887

Actions (login required)

Перегляд елементу