Abstract
This paper presents Libras SignWriting Handshape (LSWH100), a new handshape dataset focused on Sign Language Recognition. The dataset includes 144,000 synthetic images of a realistic human hand, covering 100 distinct handshape classes used in Brazilian Sign Language (Libras). Handshapes are named using the convention from SignWriting, a writing system for sign languages. The dataset contains annotations for classification, detection, segmentation, depth estimation, and 3D hand keypoints. Images include indoor and outdoor scenes during different times of day, centered on a single hand that can change size, 3D rotation, and skin tone. We generated these images using Blender, a free and open-source 3D creation software. This is a challenging dataset that can be further explored. With a focus on sign language, this dataset has the potential to advance sign language recognition systems, positively impacting those who rely on sign language for communication.