Unconstrained Face Detection Dataset (UFDD) aims to fuel further research in unconstrained face detection. Source: Pushing the Limits of Unconstrained Face Detection: a Challenge Dataset and Baseline Results Image Source: https://ufdd.info/

A new face annotation dataset with balanced distribution between genders and ethnic origins. Source: SensitiveNets: Learning Agnostic Representations with Application to Face Images

Real-World Masked Face Dataset (RMFD) is a large dataset for masked face detection. Source: https://github.com/X-zhangyang/Real-World-Masked-Face-Dataset Image Source: https://github.com/X-zhangyang/Real-World-Masked-Face-Dataset

Although deep face recognition has achieved impressive results in recent years, there is increasing controversy regarding racial and gender bias of the models, questioning their trustworthiness and deployment into sensitive scenarios. DemogPairs is a validation set with 10.8K facial images and 58.3M identity verification pairs, distributed in demographically-balanced folds of Asian, Black and White females and males. We also propose a benchmark of experiments using DemogPairs over state-of-the-art deep face recognition models in order to analyze their cross-demographic behavior and potential demographic biases (see figure below).

QMUL-SurvFace is a surveillance face recognition benchmark that contains 463,507 face images of 15,573 distinct identities captured in real-world uncooperative surveillance scenes over wide space and time. Source: Surveillance Face Recognition Challenge

The USF Human ID Gait Challenge Dataset is a dataset of videos for gait recognition. It has videos from 122 subjects in up to 32 possible combinations of variations in factors. Source: http://www.eng.usf.edu/cvprg/Gait_Data.html

The High-Quality Wide Multi-Channel Attack database (HQ-WMCA) database consists of 2904 short multi-modal video recordings of both bona-fide and presentation attacks. There are 555 bonafide presentations from 51 participants and the remaining 2349 are presentation attacks. The data is recorded from several channels including color, depth, thermal, infrared (spectra), and short-wave infrared (spectra).

The iCartoonFace dataset is a large-scale dataset that can be used for two different tasks: cartoon face detection and cartoon face recognition. Source: https://github.com/luxiangju-PersonAI/iCartoonFace Image Source: https://github.com/luxiangju-PersonAI/iCartoonFace

3DFAW contains 23k images with 66 3D face keypoint annotations. Source: Unsupervised Learning of Probably Symmetric Deformable 3D Objects from Images in the Wild Image Source: http://mhug.disi.unitn.it/workshop/3dfaw/

Acted Facial Expressions In The Wild (AFEW) is a dynamic temporal facial expressions data corpus consisting of close to real world environment extracted from movie

A large-scale, hierarchical annotated dataset of animal faces, featuring 21.9K faces from 334 diverse species and 21 animal orders across biological taxonomy. These faces are captured `in-the-wild' conditions and are consistently annotated with 9 landmarks on key facial features. The proposed dataset is structured and scalable by design; its development underwent four systematic stages involving rigorous, manual annotation effort of over 6K man-hours. Source: AnimalWeb: A Large-Scale Hierarchical Dataset of Annotated Animal Faces

FERG is a database of cartoon characters with annotated facial expressions containing 55,769 annotated face images of six characters. The images for each character are grouped into 7 types of cardinal expressions, viz. anger, disgust, fear, joy, neutral, sadness and surprise. Source: VGAN-Based Image Representation Learningfor Privacy-Preserving Facial Expression Recognition Image Source: http://grail.cs.washington.edu/projects/deepexpr/ferg-2d-db.html

Human3.6M 3D WholeBody (H3WB) is a large scale dataset with 133 whole-body keypoint annotations on 100K images, made possible by a new multi-view pipeline. It is designed for the three new tasks : i) 3D whole-body pose lifting from 2D complete whole-body pose, ii) 3D whole-body pose lifting from 2D incomplete whole-body pose, iii) 3D whole-body pose estimation from a single RGB image. Source: H3WB: Human3.6M 3D WholeBody Dataset and Benchmark Image Source: https://arxiv.org/pdf/2211.15692v1.pdf

MeGlass is an eyeglass dataset originally designed for eyeglass face recognition evaluation. All the face images are selected and cleaned from MegaFace. Each identity has at least two face images with eyeglass and two face images without eyeglass. It contains 47,817 images from 1,710 different identities. Source: https://github.com/cleardusk/MeGlass Image Source: https://github.com/cleardusk/MeGlass

The Oulu-NPU face presentation attack detection database consists of 4950 real access and attack videos. These videos were recorded using the front cameras of six mobile devices (Samsung Galaxy S6 edge, HTC Desire EYE, MEIZU X5, ASUS Zenfone Selfie, Sony XPERIA C5 Ultra Dual and OPPO N3) in three sessions with different illumination conditions and background scenes. The presentation attack types considered in the OULU-NPU database are print and video-replay. The 2D face artefacts were created using two printers and two display devices. The videos of the 55 subjects are divided into three subject-disjoint subsets for training, development and testing. Four test protocols are used to evaluate the generalization capability of face PAD methods across three covariates: unknown environmental conditions (namely illumination and background scene), acquisition devices and presentation attack instruments (PAI). Each of the four unambiguously defined evaluation protocols introduces at least one previously unseen condition to the test set, which enables a fair comparison on the generalization capabilities between new and existing approaches. Image Source: https://www.researchgate.net/profile/Neil-Robertson/publication/333834759/figure/fig5/AS:897964780302339@1591102895306/Samples-from-the-OULU-NPU-database-From-top-to-bottom-is-the-three-sessions-with_W640.jpg

The SAMM Long Videos dataset consists of 147 long videos with 343 macro-expressions and 159 micro-expressions. The dataset is FACS-coded with detailed Action Units. Source: SAMM Long Videos: A Spontaneous Facial Micro- and Macro-Expressions Dataset

Biwi Kinect Head Pose is a challenging dataset mainly inspired by the automotive setup. It is acquired with the Microsoft Kinect sensor, a structured IR light device. It contains about 15k frame, with RGB. (640 × 480) and depth maps (640 × 480). Twenty subjects have been involved in the recordings: four of them were recorded twice, for a total of 24 sequences. The ground truth of yaw, pitch and roll angles is reported together with the head center and the calibration matrix. Source: POSEidon: Face-from-Depth for Driver Pose Estimation

The Chinese Academy of Sciences Micro-Expression dataset (CASME II) consists of 255 videos, elicited from 26 participants. The videos are recorded using Point Gray GRAS-03K2C camera which has a frame rate of 200fps. The average video length is 0.34s, equivalent to 68 frames. Each video’s emotion label is annotated by two coders, where the reliability is 0.846. All the images are cropped to 170×140 pixels. The ground-truth information provided by the database include the emotion state, the action unit, the onset, apex and offset frame indices. The videos are grouped into seven categories: others (99 videos), disgust (63 videos), happiness (32 videos), repression (27 videos), surprise (25 videos), sadness (7 videos) and fear (2 videos). Source: OFF-ApexNet on Micro-expression Recognition System

Contains over 11000 images of 1000 identities with different types of disguise accessories. The dataset is collected from the Internet, resulting in unconstrained face images similar to real world settings. Source: Recognizing Disguised Faces in the Wild

MegaAge is a large dataset that consists of 41,941 faces annotated with age posterior distributions.

300 Videos in the Wild (300-VW) is a dataset for evaluating facial landmark tracking algorithms in the wild. The dataset authors collected a large number of long facial videos recorded in the wild. Each video has duration of ~1 minute (at 25-30 fps). All frames have been annotated with regards to the same mark-up (i.e. set of facial landmarks) used in the 300 W competition as well (a total of 68 landmarks). The dataset includes 114 videos (circa 1 min each).

Drone Surveillance of Faces, is a large-scale drone dataset intended to facilitate research for face recognition using drones. Data has been captured at two outdoor locations: (i) at the ground level, where subjects are asked to walk in a park-like environment, and (ii) at the terrace of a building. For each location and surveillance scenario, data is captured twice: (i) during the morning and (ii) during the evening, before sunset. For each combination of location, surveillance use case, and time of capture, there exist 25 videos featuring 58 subjects. Each subject belongs to the age bracket of (18, 40) years, and each video contains subjects appearing in groups of 2-3. For a particular setting combination, one group appears in only one video. Since there exist eight combinations of location, surveillance scenario, and time of capture, each group of subjects occur in eight videos, thus resulting in a total of 200 videos. The subjects and pairings remain consistent across different settings. The directory structure represents the combination of location, use case and time of capture.

FFHQ-Aging is a Dataset of human faces designed for benchmarking age transformation algorithms as well as many other possible vision tasks. This dataset is an extention of the NVIDIA FFHQ dataset, on top of the 70,000 original FFHQ images, it also contains the following information for each image: * Gender information (male/female with confidence score) * Age group information (10 classes with confidence score) * Head pose (pitch, roll & yaw) * Glasses type (none, normal or dark) * Eye occlusion score (0-100, different score for each eye) * Full semantic map (19 classes, based on CelebAMask-HQ labels) Source: https://github.com/royorel/FFHQ-Aging-Dataset Image Source: https://github.com/royorel/FFHQ-Aging-Dataset

The German Lipreading dataset consists of 250,000 publicly available videos of the faces of speakers of the Hessian Parliament, which was processed for word-level lip reading using an automatic pipeline. The format is similar to that of the English language Lip Reading in the Wild (LRW) dataset, with each H264-compressed MPEG-4 video encoding one word of interest in a context of 1.16 seconds duration, which yields compatibility for studying transfer learning between both datasets. Choosing video material based on naturally spoken language in a natural environment ensures more robust results for real-world applications than artificially generated datasets with as little noise as possible. The 500 different spoken words ranging between 4-18 characters in length each have 500 instances and separate MPEG-4 audio- and text metadata-files, originating from 1018 parliamentary sessions. Additionally, the complete TextGrid files containing the segmentation information of those sessions are also included. The size of the uncompressed dataset is 15GB.

iQIYI-VID dataset, which comprises video clips from iQIYI variety shows, films, and television dramas. The whole dataset contains 500,000 videos clips of 5,000 celebrities. The length of each video is 1~30 seconds. Source: iQIYI-VID

MCXFace is a heterogeneous face recognition dataset consisting of multi-channel image samples for 51 subjects. For each subject color (RGB), thermal, near-infrared (850 nm), short-wave infrared (1300 nm), Depth, Stereo depth, and depth estimated from RGB images are available. Overall 7406 images together with landmark annotations and standard protocols are available in this dataset. The Multi-Channel Heterogeneous Face Recognition dataset (MCXFace) is derived from the HQ-WMCA dataset (https://www.idiap.ch/en/dataset/hq-wmca). The MCXFace dataset contains images of 51 subjects collected in different channels under three different sessions and various illumination conditions. The channels available are color (RGB), thermal, near-infrared (850 nm), short-wave infrared (1300 nm), Depth, Stereo depth, and depth estimated from RGB images using the 3DDFA method. All the channels are registered spatially and temporally across all the modalities. In the MCXFace dataset, only bonafide samples are present. Further, the files are divided into train and dev sets with a disjoint set of identities to make experiments in different homogeneous and heterogeneous settings possible. For each of the protocols, we have created five different folds, by randomly dividing the subjects into train and dev partitions. In addition to the images, annotations for left and right eye centers for all the images are also provided in a JSON file. Specifically for each image, annotations for the left eye, right eye, mouth left corner, mouth right corner, nose, top left, and top right are provided. Each image file is stored as a ``.jpg'' file with a resolution of 1920x1200.

A multimodal database for eye blink detection and attention level estimation. Source: mEBAL: A Multimodal Database for Eye Blink Detection and Attention Level Estimation

Contains video clips shot with modern high-resolution mobile cameras, with strong projective distortions and with low lighting conditions. Source: MIDV-2019: Challenges of the modern mobile-based document OCR

The Masked LFW (MLFW), based on Cross-Age LFW (CALFW) database, is built using a simple but effective tool that generates masked faces from unmasked faces automatically. Image source: https://arxiv.org/pdf/2109.05804v1.pdf

Data Description The training data contains twelve-lead ECGs. The validation and test data contains twelve-lead, six-lead, four-lead, three-lead, and two-lead ECGs: 1. Twelve leads: I, II, III, aVR, aVL, aVF, V1, V2, V3, V4, V5, V6 2. Six leads: I, II, III, aVR, aVL, aVF 3. Four leads: I, II, III, V2 4. Three leads: I, II, V2 5. Two leads: I, II Each ECG recording has one or more labels that describe cardiac abnormalities (and/or a normal sinus rhythm). We mapped the labels for each recording to SNOMED-CT codes. The lists of scored labels and unscored labels are given with the evaluation code; see the scoring section for details. # Data Sources The Challenge data include recordings from last year’s Challenge and many new recordings for this year’s Challenge: 1. CPSC Database and CPSC-Extra Database 2. INCART Database 3. PTB and PTB-XL Database 4. The Georgia 12-lead ECG Challenge (G12EC) Database 5. Augmented Undisclosed Database 6. Chapman-Shaoxing and Ningbo Database 7. The University of Michigan (UMich) Database The Challenge data include annotated twelve-lead ECG recordings from six sources in four countries across three continents. These databases include over 100,000 twelve-lead ECG recordings with over 88,000 ECGs shared publicly as training data, 6,630 ECGs retained privately as validation data, and 36,266 ECGs retained privately as test data. - The first source is the China Physiological Signal Challenge in 2018 (CPSC 2018), which was held during the 7th International Conference on Biomedical Engineering and Biotechnology in Nanjing, China. This source contains two databases: the data from CPSC 2018 (the CPSC Database) and unused data from CPSC 2018 (the CPSC-Extra Database). Together, these databases contain 13,256 ECGs (10,330 ECGs shared as training data, 1,463 retained as validation data, and 1,463 retained as test data). We shared the training set and an unused dataset from CPSC 2018 as training data, and we split the test set from CPSC 2018 into validation and test sets. Each recording is between 6 and 144 seconds long with a sampling frequency of 500 Hz. - The second source is the St Petersburg INCART 12-lead Arrhythmia Database. This source contains 74 annotated ECGs (all shared as training data) extracted from 32 Holter monitor recordings. Each recording is 30 minutes long with a sampling frequency of 257 Hz. - The third source is the Physikalisch-Technische Bundesanstalt (PTB) and includes two public datasets: the PTB and the PTB-XL databases. The source contains 22,353 ECGs (all shared as training data). Each recording is between 10 and 120 seconds long with a sampling frequency of either 500 or 1,000 Hz. - The fourth source is a Georgia database which represents a unique demographic of the Southeastern United States. This source contains 20,672 ECGs (10,344 ECGs shared as training data, 5,167 retained as validation data, and 5,161 retained as test data). Each recording is between 5 and 10 seconds long with a sampling frequency of 500 Hz. - The fifth source is an undisclosed American database that is geographically distinct from the Georgia database. This source contains 10,000 ECGs (all retained as test data). - The sixth source is the Chapman University, Shaoxing People’s Hospital (Chapman-Shaoxing) and Ningbo First Hospital (Ningbo) database. This source contains 45,152 ECGS (all shared as training data). Each recording is 10 seconds long with a sampling frequency of 500 Hz. - The seventh source is UMich Database from the University of Michigan. This source contains 19,642 ECGs (all retained as test data). Each recording is 10 seconds long with a sampling frequency of either 250 Hz or 500 Hz. Like other real-world datasets, different databases may have different proportions of cardiac abnormalities, but all of the labels in the validation or test data are represented in the training data. Moreover, while this is a curated dataset, some of the data and labels are likely to have errors, and an important part of the Challenge is to work out these issues. In particular, some of the databases have human-overread machine labels with single or multiple human readers, so the quality of the labels varies between databases. You can find more information about the label mappings of the Challenge training data in this table. The six-lead, four-lead, three-lead, and two-lead validation data are reduced-lead versions of the twelve-lead validation data: the same recordings with the same header data but only with signal data for the relevant leads. We are not planning to release the test data at any point, including after the end of the Challenge. Requests for the test data will not receive a response. We do not release test data to prevent overfitting on the test data and claims or publications of inflated performances. We will entertain requests to run code on the test data after the Challenge on a limited basis based on publication necessity and capacity. (The Challenge is largely staged by volunteers.) # Data Format All data was formatted in WFDB format. Each ECG recording uses a binary MATLAB v4 file (see page 27) for the ECG signal data and a plain text file in WFDB header format for the recording and patient attributes, including the diagnosis, i.e., the labels for the recording. The binary files can be read using the load function in MATLAB and the scipy.io.loadmat function in Python; see our MATLAB and Python example code for working examples. The first line of the header provides information about the total number of leads and the total number of samples or time points per lead, the following lines describe how each lead was encoded, and the last lines provide information on the demographics and diagnosis of the patient. For example, a header file A0001.hea may have the following contents: A0001 12 500 7500 05-Feb-2020 11:39:16 A0001.mat 16+24 1000/mV 16 0 28 -1716 0 I A0001.mat 16+24 1000/mV 16 0 7 2029 0 II A0001.mat 16+24 1000/mV 16 0 -21 3745 0 III A0001.mat 16+24 1000/mV 16 0 -17 3680 0 aVR A0001.mat 16+24 1000/mV 16 0 24 -2664 0 aVL A0001.mat 16+24 1000/mV 16 0 -7 -1499 0 aVF A0001.mat 16+24 1000/mV 16 0 -290 390 0 V1 A0001.mat 16+24 1000/mV 16 0 -204 157 0 V2 A0001.mat 16+24 1000/mV 16 0 -96 -2555 0 V3 A0001.mat 16+24 1000/mV 16 0 -112 49 0 V4 A0001.mat 16+24 1000/mV 16 0 -596 -321 0 V5 A0001.mat 16+24 1000/mV 16 0 -16 -3112 0 V6 #Age: 74 #Sex: Male #Dx: 426783006 #Rx: Unknown #Hx: Unknown #Sx: Unknown From the first line of the file, we see that the recording number is A0001, and the recording file is A0001.mat. The recording has 12 leads, each recorded at a 500 Hz sampling frequency, and contains 7500 samples. From the next 12 lines of the file (one for each lead), we see that each signal was written at 16 bits with an offset of 24 bits, the floating point number (analog-to-digital converter (ADC) units per physical unit) is 1000/mV, the resolution of the analog-to-digital converter (ADC) used to digitize the signal is 16 bits, and the baseline value corresponding to 0 physical units is 0. The first value of the signal (-1716, etc.), the checksum (0, etc.), and the lead name (I, etc.) are the last three entries of each of these lines. From the final 6 lines, we see that the patient is a 74-year-old male with a diagnosis (Dx) of 426783006, which is the SNOMED-CT code for sinus rhythm. The medical prescription (Rx), history (Hx), and symptom or surgery (Sx) are unknown. Please visit WFDB header format for more information on the header file and variables.

Aims to facilitate research in caricature recognition. All the caricatures and face images were collected from the Web. Compared with two existing datasets, this dataset is much more challenging, with a much greater number of available images, artistic styles and larger intra-personal variations. Source: WebCaricature: a benchmark for caricature recognition

Large, multimodal biometric dataset: It contains still images and videos of over 1,000 people captured at various ranges (up to 1,000 meters) and elevations (up to 400 meters) using a diverse set of cameras (commercial, military-grade, specialized). Focus on challenging scenarios: Designed to address limitations of existing datasets, it includes data with extreme poses, low resolutions, and atmospheric distortions to push the boundaries of biometric recognition algorithms. Rich annotations and diversity: Each data point is accompanied by detailed metadata (sensor details, weather conditions) and manual annotations (bounding boxes, yaw/pitch angles) for face and whole-body regions. The dataset also ensures demographic diversity for robust algorithm development.

Jiyoung Lee, Seungryong Kim, Sunok Kim, Jungin Park, Kwanghoon Sohn; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 10143-10152

FFHQ-UV is a large-scale facial UV-texture dataset that contains over 50,000 high-quality texture UV-maps with even illuminations, neutral expressions, and cleaned facial regions, which are desired characteristics for rendering realistic 3D face models under different lighting conditions. The dataset is derived from FFHQ and preserves the most variations in FFHQ. Source: FFHQ-UV: Normalized Facial UV-Texture Dataset for 3D Face Reconstruction Image Source: https://arxiv.org/pdf/2211.13874v1.pdf

FS2K is a high-quality Facial Sketch Synthesis (FSS). It consists of 2,104 image-sketch pairs spanning three types of sketch styles, image backgrounds, lighting conditions, skin colors, and facial attributes. FS2K differs from previous FSS datasets in difficulty, diversity, and scalability, and should thus facilitate the progress of FSS research.

HOD is a dataset for 3D object reconstruction which contains 35 objects, divided into two subsets named Sculptures and Daily Objects. The Sculptures has five human sculptures with complex geometries and pure white textures. The Daily Objects consists of 30 daily objects with various shapes and appearances. All of the Sculptures and nine of the Daily Objects are paired with high-fidelity scanned meshes as ground truth geometries for evaluation. Source: Reconstructing Hand-Held Objects from Monocular Video Image Source: https://arxiv.org/pdf/2211.16835v1.pdf