Leakage-Aware BISINDO Alphabet Classification with Hand-Region DINOv2 Embeddings
Indonesian Journal of Data and Science · 17.07.2026 · 7 (2) · 389-403
DOI: 10.56705/ijodas.v7i2.528 ↗Kurz gefasst
Fachartikel in der Zeitschrift Indonesian Journal of Data and Science, online erschienen am 17.07.2026. Verfasst von Lilis Nur Hayati und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).
Abstract (Originalsprache)
Static sign-alphabet classifiers can be evaluated on image partitions in which background, actor, or capture context is visually correlated across folds. This study evaluates 26-class BISINDO alphabet classification from annotated hand regions while making such context disjoint between development and final testing. Of 19,753 source images in a public corpus, 19,746 had matched annotations and yielded 19,748 labelled hand instances. A similarity screen identified 961 high-similarity source-train/source-test visual-context candidates; therefore, a deterministic stratified group-disjoint 7:1:1 partition was used (14,107/2,821/2,820 instances). A grayscale PCA-plus-logistic-regression baseline was compared with frozen DINOv2-small embeddings and class-balanced logistic regression under raw, fixed-gamma, and adaptive-gamma input conditions. Raw DINOv2 was selected exclusively by validation macro-F1. On the sealed 2,820-instance test set, the primary raw model achieved macro-F1 0.898 (95% bootstrap CI 0.885–0.910), balanced accuracy 0.898, accuracy 0.903, expected calibration error 0.030, and multiclass Brier score 0.143. The pre-specified adaptive-gamma condition attained macro-F1 0.904 after the selection decision, but it did not replace the validation-selected primary model. Performance declined under synthetic brightening and darkening, and several visually similar alphabet classes remained challenging. The results establish a reproducible hand-region benchmark under visual-context control; they do not establish detector, signer-independent, fairness, or continuous-sign performance.
Zitieren
APA
Hayati, L. N., & Indra, D. (2026). Leakage-Aware BISINDO Alphabet Classification with Hand-Region DINOv2 Embeddings. Indonesian Journal of Data and Science, 7(2), 389-403. https://doi.org/10.56705/ijodas.v7i2.528
Vancouver
Hayati LN, Indra D. Leakage-Aware BISINDO Alphabet Classification with Hand-Region DINOv2 Embeddings. Indonesian Journal of Data and Science. 2026;7(2):389-403. doi:10.56705/ijodas.v7i2.528
BibTeX
@article{Hayati2026,
author = {Hayati, Lilis Nur and Indra, Dolly},
title = {Leakage-Aware BISINDO Alphabet Classification with Hand-Region DINOv2 Embeddings},
journal = {Indonesian Journal of Data and Science},
year = {2026},
volume = {7},
number = {2},
pages = {389--403},
doi = {10.56705/ijodas.v7i2.528}
}
EUVIMED