Word-Level BISINDO: A Novel Video Indonesian Sign Language Dataset and Baseline Methods

The Deaf community in Indonesia, which constitutes 3-4% of the global deaf population, primarily uses Bahasa Isyarat Indonesia (BISINDO), a sign language distinct from Sistem Isyarat Bahasa Indonesia (SIBI). To address the lack of publicly available datasets for BISINDO recognition, this study introduces the Word-Level BISINDO dataset. This dataset contains 1600 RGB videos of 32 isolated signs performed by 5 signers from Banten. Transformer-based models—SPOTER, Siformer, and MViTv2—were implemented and compared using this dataset. Experimental results show that MViTv2 excelled in the signer- dependent (SD) setting with a top-1 accuracy of 97.08%, while Siformer performed best in the signer-independent (SI) setting in 3 out of 4 experiments. This highlights both the robustness of the WL-BISINDO and the strong generalization capability of pose- based models. The WL-BISINDO dataset serves as a foundational resource for research on BISINDO recognition and has the potential to support the development of inclusive sign language technologies in Indonesia.