saluslab/DUET

Dataset Overview

    This repository introduces a multi-modal dataset, Dyadic User Engagement dataseT (DUET), which contains 12 two-person—or dyadic—activities. Theses activities are adopted from The Repertoire of Nonverbal Behavior: Categories, Origins, Usage, and Coding by Paul Ekman et al, which allows us to distill the semantics embedded in bodily movements. Besides increasing the number, diversity, and quality of dyadic datasets, contextualizing human activities has been proven to improve the performance of human activity recognition (HAR) tasks, as well as benefit downstream applications, such as autonomous vehicles, smart homes, healthcare, and many more. The dataset is collected via Microsoft Azure Kinect v2 and constitutes of 14,400 samples, all of which come with 4 modalities: RGB, depth, infrared (IR), and 3D skeleton joints. The following sections detail the folder structure used to categorize our data, sample frames, and the specifications of Microsoft Azure Kinect v2.

Data Collection and Management

Data modalities and data format

     For the data collection, we use the high-quality and multimodal Azure Kinect, equipped with an RGB camera, a depth sensor, and an IR sensor. These sensors all operate at 30 frames per second (FPS) for three seconds for each video sample, yielding 91 frames per sample. The specification of each data format varies depending on the conventions commonly used in the research community: each RGB frame is captured with a resolution of 1,920x1,080 and is stored in a .jpeg format. We record depth and IR sequences with a resolution of 640x576 and store them as 24-bit .png files. The skeleton joints of every sample video are stored in their corresponding .csv files. Each file contains a 91x193 array, where each row represents a frame, and each column holds information related to that frame. The first column records the timestamp of the frame, and the following 96 columns capture the x, y, and z coordinates of 32 joints of one subject (as illustrated in Figure 1), measured as the distance (in millimeters) from the joint to the camera. For instance, the first three columns record the x, y, and z values of the first joint. The order of the joints follows the joint index in Azure Kinect Body Tracking Joints. The last 96 columns record the 32 joints of the other object.

Figure 1. 32 skeleton joints of a subject extracted using the Azure Kinect software development kit (SDK).

Data acquisistion arrangement

    After selecting the Azure Kinect as the multimodal sensing module, a setup for housing the sensor was needed to guarantee consistency throughout the experiment. We built a sensing module, illustrated in Figure 2, that situates the Azure Kinect 84 inches above the ground and tilts it 37° forward to capture the interactions with a full field of view and minimal occlusions.

Figure 2. On the left, we have the bird's-eye view of the testbed configuration, whereas on the right is the sensing module used across the experiment.

    Another important aspect of the experiment is the testbeds. Three locations across a US university campus are selected to carry out the experiment. As shown in Figure 3, these include an open indoor space, a confined indoor space, and an outdoor space. These three locations are chosen (1) to enrich the variety of backgrounds and (2) investigate the effects the ambient environment imposes on the sensors. One constraint of HAR datasets is the scarcity of diverse backgrounds, which can lead to overfitting to background noise for deep learning models. The experiment is carried out at three distinct locations to improve the generalizability of background noise. We also recognize that a contextualizable dataset should be suitable for a wide range of environments (e.g., parks, schools, nursing facilities, smart homes). Collecting our dataset at different locations–especially outdoors–encourages the exploration of the direct and indirect effects the ambient environment imposes on the sensors and algorithms.

Figure 3. Data collection locations include, starting from the left to right, an open indoor space, a confined indoor space, and an open outdoor space.

    Since the experiment is carried out at three locations, there is a need to ensure the collection process is repeatable. Towards this end, we designed a testbed arrangement, shown in Figure 2, that was used across all three environments. In the testbed, volunteers are asked to perform each interaction for 40 repetitions in a rectangular area taped to the ground. After each repetition, a beep would sound, instructing the subjects to rotate either clockwise or counterclockwise and proceed to the next repetition. This novel technique collects data on the interactions from a wide array of perspectives with respect to the camera, diversifying the way interactions are captured and ameliorating the perspective invariance quality of deep learning algorithms.

Subjects

    A total of 15 male and eight female subjects participated in the experiments. The subjects were randomly paired to perform actions across the three locations. The subjects' ages range from 23 to 42 years old with a mean of 27 years old and standard deviation of 4.01 years. The subjects' heights range from 165.1cm to 185.4cm with a mean of 172.7cm and standard deviation of 8.46cm. The subjects' weights range from 55kg to 93kg with a mean of 69kg and standard deviation of 10.1kg.

Folder structure

    In this repository, we have 14,400 samples that comprise RGB, depth, IR, and 3D skeleton joints, which can be very complicated. To provide simple access for users, we have organized our data into a folder structure, as shown in Figure 5. The folder structure comprises four layers: (1) modality, (2) location combination, interaction label, and subject, (3) timestamps, and (4) image or csv files. Traversing through this structure, we first classify the files based on their modality, including RGB, depth, IR, and 3D skeleton joints. The next layer classifies the location, interaction label, and subject using six-digit codes, LLIISS. Here, LL stands for the location, which can be CM for the indoor open space, CC for the indoor confined space, or CL for the outdoor space. Next, II denotes numbers ranging from 1–12, where each number corresponds to the enumeration of activities listed in the table below. Last, SS identifies the subject pairs ranging from 1–10. It is worth noting that the same subject pair number in different locations does not represent the same pair. In fact, only CCII02 and CLII07, CCII01 and CMII10, and CCII03 and CMII05 share the same subject pairs, respectively. Also, as previously mentioned, we ask each pair of subjects to repeat an interaction for 40 times, all of which are recorded in the same video. To temporally segment each clip, we classify each time window by the start and finish time marks. For example, a folder named 40800222_43800211 contains a recording starting from 40800222 and ending at 43800211. The clock, which generates the timestamps in milliseconds, begins once the Azure Kinect is connected. Every timestamp folder stores the clip of the corresponding time window, frame by frame, in which all frames are chronologically ordered by numbers ranging from 0–90.

Figure 4. The data folder structure for our dataset, which is designed for easy user access. Here, RGB, depth, and IR modalities share an identidcal hierarchy, while 3D skeleton joint folders store all 3D coordinates of a sample clip in a single .csv file.

Label ID	Dyadic interaction
1	Waving in
2	Thumbs up
3	Waving
4	Painting
5	Showing measurements
6	Nodding
7	Drawing circles in the air
8	Holding palms out
9	Twirling or scratching hair
10	Laughing
11	Arm crossing
12	Hugging

Table 1. Activity labels and their corresponding interactions.

Sample frames

    Sample frames are provided in Figure 6 to visualize the differences between different modalities, each of which possess different strengths and weaknesses. RGB frames capture information-rich features like interaction, location, and characteristic features of subjects, which are informative but fail to prioritize user privacy. However, since RGB frames compress the 3D world into a 2D plane, they often suffer from occlusion and variation in perspective. On the other hand, 3D skeleton joints reveal the placement of each joint in the 3D space. The additional dimension gives 3D skeleton joints a desirable perspective-invariant characteristic. Besides the 3D position of each joint, no further information indicative of the subject is conspicuous, prioritizing the preservation of privacy. This feature is preferred by human-centered applications, such as smart homes, CPSIS, and elder care management. Overall, the juxtaposition of different modalities exemplifies the inversely proportional relationship between privacy and value of information---the more information a modality carries, the less user privacy it typically protects. We provide four modalities in our dataset that span this full spectrum to encourage both the exploration of a single modality and the fusion of multiple modalities to strike a balance between privacy preservation and value of information.

Figure 5. Sample data of 12 interactions. Modalities presented are, from top row to bottom row: RGB, IR, depth, and 3D skeleton joints. The 12 interactions are, from left to right: waving in, thumbs up, waving, pointing, showing measurements, nodding, drawing circles in the air, holding palms out, twirling or scratching hair, laughing, arm crossing, and hugging.

Cross-location and cross-subject evaluations

One of the motivations for creating DUET is to encourage the research community to study HAR in the context of dyadic, contextualizable interactions. Hence, there is a need to provide a baseline training and test data split for algorithms to evaluate their performance. In addition to the basic cross-subject evaluation, we include a cross-location evaluation. We recognize that applications leveraging dyadic, contextualizable interactions might occur in various locations, both indoor and outdoors. Therefore, we include cross-location evaluation for HAR algorithm training to ensure resilience to location variation. For the cross-subject evaluation, we use CCII05, CCII07, CLII01, CLII05, CMII06, and CMII09 for the test data, and the remainder for the training data. For cross-location evaluation, CCIISS is selected as the test data, while CLIISS and CMIISS are used as the training data.