A comprehensive dataset of static and dynamic signs for the Mexican Sign Language alphabet Authors: Irvin Hussein Lopez-Nava, Jesús Antonio Navarrete-López, Ricardo Fernando Morfín-Chávez Published: 2026 ABSTRACT AbstractThis data article presents two complementary datasets that together cover the complete 27-letter Mexican Sign Language (Lengua de Señas Mexicana, LSM) alphabet, encompassing both static and dynamic signs. The first dataset contains 279,716 JPEG images of the 21 static signs (letters A–I, L–P, R–U, W, Y) captured from 20 participants across three levels of hand-pose variation. The second dataset contains 1,200 MP4 videos of the 6 dynamic signs (letters J, K, Ñ, Q, X, Z) recorded from 20 different participants in two camera views (frontal and profile), with five repetitions per sign. Both datasets were collected under controlled conditions using a uniform green background and constant illumination, totaling approximately 8.4 GB. Each dataset includes predefined participant-based train/test partitions for signer-independent evaluation. The datasets have been used in prior work on fingerspelling recognition [3] and unified alphabet classification [4]. The data are publicly available in the Zenodo repository under a CC-BY 4.0 license.Keywords: Mexican Sign Language, LSM, Sign language recognition, Static signs, Dynamic signs, Hand gesture dataset, Fingerspelling, Image dataset, Video dataset Specifications Table SubjectComputer Science, Artificial Intelligence, Human-Computer InteractionSpecific subject areaSign language recognition; hand gesture datasets for machine learningType of dataImages (JPEG), Videos (MP4)Data collectionStatic signs were captured as video sequences using a Nikon D5500 camera at 1920 × 1080 px, 60 fps [1], with frames subsequently extracted as JPEG images. Dynamic signs were recorded as MP4 video sequences using the same Nikon D5500 camera at 1280 × 720 px, 30 fps [2]. Both datasets were collected in a controlled environment with constant illumination (softboxes and diffuser umbrellas) and a uniform green background.Data source locationInstitution: Centro de Investigación Científica y de Educación Superior de Ensenada (CICESE)City/Region: Ensenada, Baja CaliforniaCountry: MexicoData accessibilityRepository name: ZenodoStatic signs: https://doi.org/10.5281/zenodo.10067509Dynamic signs: https://doi.org/10.5281/zenodo.14689869License: Creative Commons Attribution 4.0 International (CC-BY 4.0)Related research articleR.F. Morfín-Chávez et al., Fingerspelling Recognition in Mexican Sign Language (LSM) Using Machine Learning, MICAI 2023, LNAI 14391, Springer, 2024, pp. 110–120. doi:10.1007/978-3-031-47765-2_9J.A. Navarrete-López et al., Dynamic Strategy for Recognizing the Mexican Sign Language Alphabet, MCPR 2025, LNCS 15715, Springer, 2025, pp. 113–122. doi:10.1007/978-3-031-96255-4_11 Value of the Data •While previous LSM datasets have focused on subsets of the alphabet or specific modalities—such as the 21-static-sign dataset by Ríos-Figueroa et al. [3] and the fingerspelling detection resource by Sánchez-Vicinaiz et al. [4]—the present datasets are, to our knowledge, the first publicly available resource covering the complete 27-letter LSM alphabet, integrating both static images and dynamic video recordings with multiple levels of hand-pose variation, dual camera perspectives, and predefined participant-based train/test splits. While previous LSM datasets have focused on subsets of the alphabet or specific modalities—such as the 21-static-sign dataset by Ríos-Figueroa et al. [3] and the fingerspelling detection resource by Sánchez-Vicinaiz et al. [4]—the present datasets are, to our knowledge, the first publicly available resource covering the complete 27-letter LSM alphabet, integrating both static images and dynamic video recordings with multiple levels of hand-pose variation, dual camera perspectives, and predefined participant-based train/test splits. •The controlled capture conditions (green background, constant illumination, standardized participant positioning) make the data suitable for training and benchmarking computer vision and deep learning models for sign language recognition. The controlled capture conditions (green background, constant illumination, standardized participant positioning) make the data suitable for training and benchmarking computer vision and deep learning models for sign language recognition. •The inclusion of multiple levels of hand-pose variation (Groups A, B, C) in the static dataset and two camera perspectives (frontal and profile) in the dynamic dataset enables research on model robustness to viewpoint and articulation changes. The inclusion of multiple levels of hand-pose variation (Groups A, B, C) in the static dataset and two camera perspectives (frontal and profile) in the dynamic dataset enables research on model robustness to viewpoint and articulation changes. •The participant-based train/test partitioning supports signer-independent evaluation, which is essential for developing generalizable sign language recognition systems. The participant-based train/test partitioning supports signer-independent evaluation, which is essential for developing generalizable sign language recognition systems. •The data have already been employed in published work on fingerspelling recognition [5] and unified static–dynamic alphabet classification [2], and can be further reused by researchers in accessibility technology, human-computer interaction, linguistic analysis of LSM phonology, and educational tool development for deaf and hard-of-hearing communities in Mexico. The data have already been employed in published work on fingerspelling recognition [5] and unified static–dynamic alphabet classification [2], and can be further reused by researchers in accessibility technology, human-computer interaction, linguistic analysis of LSM phonology, and educational tool development for deaf and hard-of-hearing communities in Mexico. Data Description The Mexican Sign Language (Lengua de Señas Mexicana, LSM) alphabet comprises 27 signs corresponding to the letters of the Spanish alphabet [6], [7]. Of these, 21 are static signs—represented by a fixed hand configuration (queirem)—while the remaining 6 (J, K, Ñ, Q, X, Z) are dynamic signs that additionally involve hand movement (kinema) over time [2]. The two datasets presented in this data article provide data for the complete LSM alphabet by covering both types of signs. Table 1 summarizes the main characteristics of each dataset. Table 1Overview of the two LSM alphabet datasets.Table 1 dummy alt textCharacteristicStatic datasetDynamic datasetSigns coveredA–I, L–P, R–U, W, Y (21)J, K, Ñ, Q, X, Z (6)Data formatJPEG imagesMP4 videosParticipants2020CameraNikon D5500Nikon D5500Resolution360 × 360 (hand crops)1280 × 720Frame rate60 fps30 fpsVariation groupsA (low), B (medium), C (high)Natural temporal variabilityCamera viewsFrontal onlyFrontal + profileRepetitions per sign1 video per variation group5 per viewBackgroundUniform greenUniform greenTrain/test split18/2 participants18/2 participantsTotal size∼4.8 GB∼3.6 GBDOI10.5281/zenodo.1006750910.5281/zenodo.14689869 Overview of the two LSM alphabet datasets. In both datasets, class labels are encoded in the file organization: for static signs, images are organized into subdirectories named by letter (e.g., A/, B/,..., Y/); for dynamic signs, the letter label is embedded in each video filename (e.g., S01-J-frontal-1.mp4). No frame-level annotations, temporal boundary markers, or hand landmark coordinates are included in the distributed files; the datasets consist exclusively of the raw cropped images (static) and original video recordings (dynamic). Static signs dataset The static dataset [8] contains 279,716 images of the 21 static LSM alphabet signs collected from 20 participants with varying levels of LSM familiarity [1]. Three variation groups capture different levels of hand-pose diversity: •Group A (low variation): One 3-second video per sign. Gentle rotations (∼30° max) across all degrees of freedom. Group A (low variation): One 3-second video per sign. Gentle rotations (∼30° max) across all degrees of freedom. •Group B (medium variation): Three 1-second videos per sign, each targeting one rotation axis (yaw, roll, pitch). Group B (medium variation): Three 1-second videos per sign, each targeting one rotation axis (yaw, roll, pitch). •Group C (high variation): One 3-second video per sign. Pronounced, unrestricted movements across all degrees of freedom. Group C (high variation): One 3-second video per sign. Pronounced, unrestricted movements across all degrees of freedom. Frames were extracted at the original capture rate (60 fps) and processed using MediaPipe Hands (v0.8) [9] for hand detection. For each frame where a hand was detected, the 21 hand landmarks were extracted to determine the bounding box extremes (min–max of normalized x,y coordinates), from which the geometric center was computed. A fixed 360 × 360 pixel crop was then extracted around this center. Frames where MediaPipe failed to detect a hand were not included in the dataset, resulting in a detection failure rate of approximately 6.6% [1]. The cropped hand images, with the original green background preserved, are the data distributed on Zenodo. Each group is split into train (18 participants) and test (2 participants) partitions. The directory structure is: Table 2 presents image counts per group. The number of images per participant and sign varies between 353 and 515 frames in the test set and between 3,404 and 4,447 frames in the training set, depending on the variation group and the duration of each recording. The average number of images per class across all groups is 13,319.81±394.94, indicating balanced class distribution. Table 2Image counts in the static dataset, broken down by variation group and train/test partition.Table 2 dummy alt textVariation groupTrainTestTotalGroup A (low)83,2649,43992,703Group B (medium)84,0198,84192,860Group C (high)84,8149,33994,153All groups252,09727,619279,716 Image counts in the static dataset, broken down by variation group and train/test partition. Fig. 1 shows sample images of the five vowels across the three variation groups, illustrating both the hand-pose diversity and inter-subject variability in the dataset. Examples of static images from the LSM alphabet dataset for the five vowel signs (A, E, I, O, U). For each letter, one image is shown per variation group: Group A (low variation, left), Group B (medium variation, center), and Group C (high variation, right). Images are 360 × 360 pixel hand crops with the original green background, as distributed on Zenodo. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Dynamic signs dataset The dynamic dataset [10] contains 1,200 video recordings of the 6 dynamic LSM alphabet signs (J, K, Ñ, Q, X, Z). As described in Navarrete-López et al. [2], dynamic signs have three phases: (i) initiation (queirem established), (ii) midpoint (kinema occurs), and (iii) final (hand returns to initial position, except J and Z). Data were collected from 20 participants, each performing each sign 5 times at a comfortable, self-selected pace, recorded with a Nikon D5500 at 1280 × 720 px, 30 fps [2]. Two camera perspectives were captured: frontal and profile (45°-rotated). The unconstrained execution speed and movement amplitude naturally introduced inter- and intra-subject variability in the data. Video durations ranged from 0.6 s to 3.46 s (mean 1.78±0.52 s). One video contained fewer than 30 frames but was retained to preserve completeness. The dataset is distributed as two compressed archives corresponding to each camera view: MSL-dynamic-signs-frontal-view.7z (∼2.2 GB) and MSL-dynamic-signs-profile.7z (∼1.4 GB). Each archive is organized into train (18 participants) and test (2 participants) partitions with the following structure: Video files follow the naming convention S---.mp4, where SubjectId is a numerical identifier (1–20), Letter is one of {J, K, Ñ, Q, X, Z}, View indicates the camera perspective (frontal or profile), and RepetitionNumber ranges from 1 to 5. Table 3 summarizes the video counts per view and partition. Table 3Video counts in the dynamic dataset, broken down by camera view and train/test partition.Table 3 dummy alt textCamera viewTrainTestTotalFrontal54060600Profile54060600Both views1,0801201,200 Video counts in the dynamic dataset, broken down by camera view and train/test partition. Fig. 2 shows representative frame sequences for the 6 dynamic signs from both camera perspectives. Representative frame sequences from the dynamic LSM alphabet dataset for the six dynamic signs (J, K, Ñ, Q, X, Z). For each sign, five uniformly sampled frames illustrate the temporal progression from initiation to completion across the frontal view (left block) and profile view (right block). Kinema diagrams on the left indicate the movement direction for each sign. Fig. 3 presents the complete 27-letter LSM alphabet as represented in the combined datasets, with blue arrows indicating the movement direction of the 6 dynamic signs. Complete 27-letter LSM alphabet as represented in the combined datasets. Static signs (A–I, L–P, R–U, W, Y) are shown as single hand-crop images; dynamic signs (J, K, Ñ, Q, X, Z) are represented by a key frame with blue arrows indicating the direction of hand movement (kinema). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Combined dataset summary Together, the two datasets cover the complete 27-letter LSM alphabet (21 static + 6 dynamic), comprising 279,716 images and 1,200 videos totaling approximately 8.4 GB. The data were collected from two completely disjoint groups of 20 participants each (40 unique individuals in total); no participant took part in both data collection campaigns. For a typical classification workflow, static images can be loaded directly from the directory structure (each subdirectory corresponds to one sign class), resized to the input dimensions required by the model, and optionally normalized. For dynamic signs, each MP4 video can be decoded into a frame sequence and either sampled at uniform intervals or processed with a temporal model. The green background may be removed via chroma keying or retained as-is, depending on the application. The predefined train/test split should be preserved to ensure signer-independent evaluation. Experimental Design, Materials and Methods Participants Two separate groups of 20 volunteers each were recruited for data collection—one group for the static signs and a different group for the dynamic signs—as the datasets were collected in independent campaigns conducted at different times. Participants were LSM enthusiasts recruited from the university community who had not undergone formal LSM training with the deaf community; however, LSM experts with strong ties to the Mexican deaf community supervised the data collection sessions to ensure sign accuracy [1]. All participants used their right hand for signing. In the static signs dataset, the gender distribution comprised 13 male and 7 female participants, whereas in the dynamic signs dataset, the distribution included 11 male and 9 female participants. Additional demographic information, such as age range, was not collected. All participants provided written informed consent, approved by the Bioethics Committee of CICESE (protocol BIOETICA_HUM_2021_07). Capture environment Both datasets were captured in a controlled indoor environment at CICESE, using the same Nikon D5500 camera but with different recording configurations for each dataset. The setup comprised [1]: •Green screen: Uniform green background to facilitate hand detection and segmentation. Green screen: Uniform green background to facilitate hand detection and segmentation. •Illumination: Two softbox lights directed at the screen (eliminating participant shadows) and two diffuser umbrellas illuminating the participant. Illumination: Two softbox lights directed at the screen (eliminating participant shadows) and two diffuser umbrellas illuminating the participant. •Positioning: Participants stood at a fixed position, extending their right hand in front of their body. Positioning: Participants stood at a fixed position, extending their right hand in front of their body. Static signs capture protocol A Nikon D5500 camera was configured at 1920 × 1080 px, 60 fps, shutter speed 1/649. Each participant performed 21 static signs following the LSM dictionary [6]. Three variation groups (A, B, C) were recorded per sign, as described in Section 1.1. After recording, all video frames were extracted and processed with MediaPipe Hands (v0.8) [9] using the following parameters: static_image_mode = False, max_num_hands = 1, min_detection_confidence = 0.5, min_tracking_confidence = 0.5. For each detected hand, the 21 landmark coordinates were used to compute the bounding box and its geometric center, around which a 360 × 360 px crop was extracted. Frames where no hand was detected were excluded from the final dataset [1]. The Python scripts implementing this preprocessing pipeline are publicly available at https://github.com/XxANTONIO73xX/lsm-alphabet-preprocessing. Dynamic signs capture protocol The same Nikon D5500 camera was used, reconfigured at 1280 × 720 px, 30 fps [2]. Each participant performed each of the 6 dynamic signs 5 times from two perspectives (frontal and profile), resulting in up to 60 videos per participant. No specific constraints were imposed on the speed or amplitude of the hand movements; participants performed each sign at a comfortable pace, which naturally introduced variability in both execution speed and the extent of movement within the signing space. The mean video duration was 1.78±0.52 s (range: 0.6–3.46 s). Data partitioning Both datasets use a participant-based 18/2 train/test split. The two groups of participants are completely disjoint (i.e., 40 unique individuals across both datasets). The same deterministic selection criterion was applied in both cases: participants with the two highest subject IDs were assigned to the test set. This non-random approach was chosen to ensure a fixed, reproducible partition across experiments and to facilitate consistent benchmarking; researchers requiring alternative splits can reassign participants using the subject identifiers embedded in the file names and directory structure. Quality control The same quality control process was applied to both datasets. Hand centering was verified using the bounding box provided by MediaPipe Hands (v0.8) [9], and a random sample of 10% of the images (static dataset) and video frames (dynamic dataset) was visually inspected to confirm correct hand detection and positioning. In the static signs dataset, MediaPipe was used as part of the processing pipeline to detect, center, and crop hand regions. It failed to detect the hand in approximately 6.6% of the extracted frames [1]. These frames were excluded during the cropping stage and are therefore not part of the released dataset; the 279,716 images distributed on Zenodo correspond exclusively to frames where a hand was successfully detected and cropped. In the dynamic signs dataset, MediaPipe was used only for quality verification purposes—not for cropping or filtering. Videos were recorded and distributed as original MP4 camera recordings (not reconstructed from individual frames). MediaPipe hand detection was applied to sampled frames to verify hand visibility, but the videos themselves were not modified based on detection outcomes. One video was identified as having fewer than 30 frames; however, no videos were removed from the publicly released dataset. All 1,200 videos, including those with occasional detection limitations, remain part of the published data. Limitations Several limitations of the datasets should be noted. Both datasets were collected under controlled studio conditions (uniform green background, constant illumination), which may limit the direct applicability of trained models to unconstrained real-world environments; deploying such models in practice would likely require domain adaptation techniques, data augmentation with varied backgrounds and lighting, or fine-tuning on in-the-wild data. All participants were right-handed, so the datasets do not capture left-handed signing variability. The participant sample (20 per dataset, 40 unique total) is comparable in size to other publicly available sign language alphabet datasets [3], [4], but does not capture the full demographic diversity of LSM signers; age information was not collected. Finally, the datasets contain only raw images and videos with directory-level and filename-level labels; no frame-level annotations, temporal segmentation boundaries, or extracted hand landmarks are included in the distributed files. Ethics Statement The data collection protocols for both datasets were approved by the Bioethics Committee of CICESE (Comité de Bioética del CICESE, protocol BIOETICA_HUM_2021_07). All participants provided written informed consent prior to data collection, including consent for the public distribution of images and videos of their hands. CRediT Author Statement I.H. Lopez-Nava: Conceptualization, Methodology, Validation, Formal analysis, Resources, Writing – original draft, Visualization, Supervision, Project administration, Funding acquisition. J.A. Navarrete-López: Software, Validation, Investigation, Data curation (dynamic dataset), Writing – review & editing. R.F. Morfín-Chávez: Software, Validation, Investigation, Data curation (static dataset). Acknowledgements This work was supported by the Secretaría de Ciencia, Humanidades, Tecnología e Innovación (SECIHTI) through the research project IH-2025-1-688 (Convocatoria Investigación Humanística 2025), and through Master’s degree grants 1294009 (J.A.N.-L.) and 1107389 (R.F.M.-C.). The authors thank the Asociación Regional de Sordos Ensenadenses and all the volunteers who participated in the data collection. Declaration of Competing Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this data article. Footnotes Supplementary material associated with this article can be found, in the online version, at 10.1016/j.dib.2026.112887 Contributor Information Irvin Hussein Lopez-Nava, Email: hussein@cicese.edu.mx. Jesús Antonio Navarrete-López, Email: jnavarrete@cicese.edu.mx. Ricardo Fernando Morfín-Chávez, Email: morfinrf@cicese.edu.mx. Supplementary materials Supplementary Data S1 Supplementary Raw Research Data. This is open data under the CC BY license http://creativecommons.org/licenses/by/4.0/ Supplementary Data S2 Supplementary Raw Research Data. This is open data under the CC BY license http://creativecommons.org/licenses/by/4.0/ Data Availability Mexican Sign Language Alphabet (static signs only): https://doi.org/10.5281/zenodo.10067509, Mexican Sign Language Alphabet (dynamic signs only): https://doi.org/10.5281/zenodo.14689869. Mexican Sign Language Alphabet (static signs only): https://doi.org/10.5281/zenodo.10067509, Mexican Sign Language Alphabet (dynamic signs only): https://doi.org/10.5281/zenodo.14689869. References Associated Data Supplementary Materials Supplementary Data S1 Supplementary Raw Research Data. This is open data under the CC BY license http://creativecommons.org/licenses/by/4.0/ Supplementary Data S2 Supplementary Raw Research Data. This is open data under the CC BY license http://creativecommons.org/licenses/by/4.0/ Data Availability Statement Mexican Sign Language Alphabet (static signs only): https://doi.org/10.5281/zenodo.10067509, Mexican Sign Language Alphabet (dynamic signs only): https://doi.org/10.5281/zenodo.14689869. Mexican Sign Language Alphabet (static signs only): https://doi.org/10.5281/zenodo.10067509, Mexican Sign Language Alphabet (dynamic signs only): https://doi.org/10.5281/zenodo.14689869.