Loading...
Loading...
Contributors:
Dataset URI:
10.13053/CyS-29-2-5119Rights:
No information provided; contact the repository owner
Facial-BSL is a multimodal RGB-D dataset of dynamic facial expressions used in Brazilian Sign Language (BSL/Libras). It contains 996 isolated facial-expression samples from 10 subjects and eight categories: Angry, Laugh, Surprise, Yawn, Cry, Full, Agree, and Empty. The expressions include intra-class and inter-class similarities, such as angry–cry, laugh–agree, and surprise–yawn, making the dataset suitable for challenging facial-expression recognition tasks. The original recordings were captured using a Microsoft Kinect V1. Each participant recorded six continuous videos of approximately 33 seconds, with an average of 15 facial expressions per video. Expressions were performed at different speeds and in different orders, producing sequences of variable duration. The samples were manually annotated with start and end times with the support of an expert in Brazilian Sign Language. Dataset composition # Category Samples # Category Samples 1 Angry 121 5 Cry 121 2 Laugh 123 6 Full 121 3 Surprise 125 7 Agree 122 4 Yawn 121 8 Empty 142 Total 996 The associated study additionally generated 4,979 augmented videos, resulting in 5,975 samples for the reported experiments. The augmented samples are not included in this release. Data organization Facial-BSL/ ├── SUBJECT_01/ │ ├── 0_Angry/ │ │ ├── 0_p1_c4_s1_1/ │ │ │ ├── Color.avi │ │ │ ├── Depth/ │ │ │ ├── landmarks_pixel.txt │ │ │ └── landmarks_realworld.txt │ │ └── ... │ ├── 1_Laugh/ │ ├── 2_Surprise/ │ ├── 3_Yawn/ │ ├── 4_Cry/ │ ├── 5_Full/ │ ├── 6_Agree/ │ └── 7_Empty/ └── SUBJECT_10/ └── ... Each sample contains: Color.avi: isolated RGB facial-expression video. A fixed face-centered crop was obtained using MediaPipe Face Landmarker while preserving the full head, neck, part of the shoulders, and natural head movement. No spatial resizing or facial alignment was applied. Videos are stored in AVI format using the MJPG codec. Depth/: original depth frames associated with the sequence. These files were copied without cropping, resizing, or re-encoding. landmarks_pixel.txt: 468 MediaPipe landmarks in pixel coordinates relative to the cropped RGB video. landmarks_realworld.txt: 468 MediaPipe landmarks represented by normalized/relative x,y,zx,y,z coordinates. The original sequence duration is preserved. The associated paper used approximately 30 frames per video and selected 290 of the 468 MediaPipe landmarks for its experiments; this release keeps the full temporal sequence and all 468 landmarks. Reading the landmark files Both landmark files contain one row per video frame and have no header. landmarks_pixel.txt contains: x0 y0 x1 y1 ... x467 y467 and can be loaded as: import numpy as np landmarks = np.loadtxt( "landmarks_pixel.txt", dtype=np.float32 ).reshape(-1, 468, 2) landmarks_realworld.txt contains: x0 y0 z0 x1 y1 z1 ... x467 y467 z467 and can be loaded as: landmarks_xyz = np.loadtxt( "landmarks_realworld.txt", dtype=np.float32 ).reshape(-1, 468, 3) The values in landmarks_realworld.txt are MediaPipe normalized/relative coordinates and must not be interpreted as Kinect metric coordinates or millimeters Reading the depth data Depth images should be loaded without automatic conversion: import cv2 depth = cv2.imread( "Depth/<depth_file>", cv2.IMREAD_UNCHANGED ) print(depth.shape, depth.dtype) The RGB video is spatially cropped, whereas the depth images retain their original spatial representation. Therefore, RGB and Depth should not be assumed to share the same pixel coordinate system without additional registration. Access and use conditions Access is restricted to academic research only. Commercial use, redistribution of the dataset or access credentials, and public display of the videos or images are prohibited. As an exception, images from Subject p3 and Subject p5 may be used in scientific publications for illustrative purposes. Images from other subjects may not be publicly reproduced. To request access, download and complete the Facial-BSL Database Agreement and send the signed form to: Lourdes Ramirez-Cernalramirec@ulima.edu.pe or Edwin Jonathan Escobedo-Cardenaseescobed@ulima.edu.pe The agreement allows the principal applicant and up to three additional researchers from the same institution to work with the dataset under the applicant's responsibility. Citation If you use this dataset, any of its modalities, or the derived facial-landmark data, please cite: Ramirez-Cerna, L., Rodriguez-Melquiades, J., Escobedo-Cardenas, E. J., Camara-Chavez, G., & Garcia-Miranda, D. (2025). Facial Expressions Recognition in Sign Language Based on a Two-Stream Swin Transformer Model Integrating RGB and Texture Map Images. Computación y Sistemas, 29(2), 773–792. https://doi.org/10.13053/CyS-29-2-5119
Keyword(s):
Brazilian Sign Language (Libras) Facial Expressions Dynamic Facial Expression Recognition Non-Manual Markers Sign Language Recognition RGB-D Dataset Multimodal Dataset Depth Imaging Facial Landmarks MediaPipe Face Landmarker Kinect Computer Vision Deep Learning Human–Computer Interaction.
MeSH Term(s):
Grants and Funding:
No Grants or Funding Data Available
Associated Publication(s):
Supplementary Data:
No Supplementary Data Available
Related Dataset(s):
No Related Dataset Data Available
The National Library of Medicine (NLM) provides information about biomedical datasets. Inclusion on an NLM Web page does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: NLM Web Policies (nih.gov)