Apoema-Libras560 - Brazilian Sign Language Multimodal Dataset

Apoema-Libras560 is a multimodal video dataset of isolated signs of Brazilian Sign Language (Libras), recorded in synchronized RGB and Depth modalities. It was collected and curated by the Núcleo de Tecnologia Assistiva — Projeto Apoema Libras (Espaço Curupira), at the Federal Institute of Education, Science and Technology of Amazonas (IFAM), to support research on sign language recognition, computer vision, multimodal learning, and assistive technologies for the Deaf community.

Dataset composition. The Apoema-Libras corpus is provided in two vocabulary scales that share the same recording apparatus, acquisition protocol, and environmental conditions. The Apoema-Libras84 subset contains 3,444 video samples covering 84 signs (6 repetitions per sign–signer pair for the training and validation signers, and 5 repetitions for the held-out test signer, one repetition having been excluded during dataset curation). The full Apoema-Libras560 set contains 23,520 video samples covering 560 signs. Both subsets are drawn from recordings of the same roster of seven signers; the two scales were segmented and partitioned independently, so a given signer may occupy different splits across the two scales, and the scales should not be combined for cross-scale transfer evaluation. Because the Kinect sensor delivers spatially and temporally registered RGB and depth streams (the depth map stored as an 8-bit image registered to the RGB frame), every sample is paired by construction, and both modalities share an identical partitioning indexed by sample IDs.

Signers. Seven adults: 4 hearing interpreters (10+ years' professional experience) and 3 deaf native signers; ages 24–38; 3 male, 4 female.

Signer-disjoint splits — Apoema-Libras560: training 13,440 (4 signers), validation 6,720 (2 signers), test 3,360 (1 signer). Apoema-Libras84: training 2,016, validation 1,008, test 420.

Acquisition. Microsoft Kinect v2 (seated signer, 1.5 m from the sensor, 1.7 m height); RGB 320×240 @ 30 fps; Depth 320×240 @ 20 fps; file format: sequential JPEG frames.

ACCESS IS RESTRICTED — for academic, non-commercial research only. The files are encrypted and are released only after a hand-signed Database Release Agreement (Standard Academic License) has been received and approved by the Owner. Access may be requested exclusively by an academic advisor/supervisor (orientador) who is formally affiliated with a recognized academic, research, or educational institution and who uses an institutional e-mail address. Students may use the dataset only under their advisor's responsibility, within the advisor's research group.

How to request access: use the "Request access" button on this record and provide your identification (Organization, Principal Researcher/advisor full name, ORCID iD, institutional e-mail, affiliation, research group). Then download, hand-sign, and e-mail the Database Release Agreement from your institutional address to apoemalibras.cmc@ifam.edu.br. Access is granted only after the signed Agreement is received and the institutional affiliation is verified.

Citation: any publication, thesis, or technical report using this dataset must cite this Zenodo record (DOI) and the reference paper once published.

Contact: apoemalibras.cmc@ifam.edu.br