BISINDO Phrase Recognition - Results, Evaluation Code, and Deposit Toolkit

Supporting material for a study on phrase-level Indonesian Sign Language (Bahasa Isyarat Indonesia, BISINDO) video recognition using a frozen convolutional backbone coupled to a mask-aware bidirectional LSTM. This release contains the complete evaluation results of the study (confusion matrices for both backbones, per-class precision/recall/F1 with test support, overall accuracy and loss, and the sampled training history), together with code that reproduces every reported metric directly from the confusion matrices and performs the paired statistical comparison between the two backbones. It also includes the toolkit that packages the cached CNN feature arrays for a planned later release. The study evaluates a 16-class phrase vocabulary over 1,243 webcam clips performed by ten native Deaf signers, comparing Inception-V3 (95.2 percent accuracy on 125 held-out clips) against ResNet50 (87.2 percent) as frozen frame-level feature extractors under an identical training protocol. This release contains no video and no personally identifying material. The raw recordings show identifiable Deaf signers and are governed separately; see docs/ETHICS.md in the archive. Code is MIT licensed; results and documentation are CC BY 4.0.