Pose Dataset Research Articles

MPE-HRNetL: A Lightweight High-Resolution Network for Multispecies Animal Pose Estimation.

Animal pose estimation is crucial for animal health assessment, species protection, and behavior analysis. It is an inevitable and unstoppable trend to apply deep learning to animal pose estimation. In many practical application scenarios, pose estimation models must be deployed on edge devices with limited resource. Therefore, it is essential to strike a balance between model complexity and accuracy. To address this issue, we propose a lightweight network model, i.e., MPE-HRNet.L, by improving Lite-HRNet. The improvements are threefold. Firstly, we improve Spatial Pyramid Pooling-Fast and apply it and the improved version to different branches. Secondly, we construct a feature extraction module based on a mixed pooling module and a dual spatial and channel attention mechanism, and take the feature extraction module as the basic module of MPE-HRNet.L. Thirdly, we introduce a feature enhancement stage to enhance important features. The experimental results on the AP-10K dataset and the Animal Pose dataset verify the effectiveness and efficiency of MPE-HRNet.L.

3D skeleton aware driver behavior recognition framework for autonomous driving system

The recognition of the driver’s behaviors inside an autonomous vehicle can effectively address emergency handling in autonomous driving and is crucial for ensuring the driver’s safety. Driver behavior recognition is a challenging task due to factors such as variations, diversities, complexities, and strong interferences in behaviors. In this paper, to realize the application in the autonomous driving scenes, a novel 3D skeleton aware behavior recognition framework is proposed to recognize various driver behaviors in autonomous driving systems. First, a 3D human pose estimation network (Pose-GTFNet) with temporal Transformer and spatial graph convolutional network (GCN) is designed to infer 3D human poses from 2D pose sequences. Second, based on the obtained 3D human pose sequences, a behavior recognition network (Beh-MSFNet) with multi-skeleton feature fusion is designed to recognize driver behaviors. In the experiments, the Pose-GTFNet and Beh-MSFNet can get the best performance compared with most state-of-the-art (SOTA) methods on the Human3.6M human pose dataset, JHMDB and SHREC action recognition dataset, respectively. In addition, the proposed driver behavior recognition framework can achieve SOTA performance on the Drive&Act and Driver-Skeleton driver behavior datasets.

3D Head Pose Estimation via Normal Maps: A Generalized Solution for Depth Image, Point Cloud, and Mesh

Head pose estimation plays a crucial role in various applications, including human–machine interaction, autonomous driving systems, and 3D reconstruction. Current methods address the problem primarily from a 2D perspective, which limits the efficient utilization of 3D information. Herein, a novel approach, called pose orientation‐aware network (POANet), which leverages normal maps for orientation information embedding, providing abundant and robust head pose information, is introduced. POANet incorporates the axial signal perception module and the rotation matrix perception module, these lightweight modules make the approach achieve state‐of‐the‐art (SOTA) performance with few computational costs. This method can directly analyze various topological 3D data without extensive preprocessing. For depth images, POANet outperforms existing methods on the Biwi Kinect head pose dataset, reducing the mean absolute error (MAE) by ≈30% compared to the SOTA methods. POANet is the first method to perform rigid head registration in a landmark‐free manner. It also incorporates few‐shot learning capabilities and achieves an MAE of about on the Headspace dataset. These features make POANet a superior alternative to traditional generalized Procrustes analysis for mesh data processing, offering enhanced convenience for human phenotype studies.

Pose‐to‐Motion: Cross‐Domain Motion Retargeting with Pose Prior

AbstractCreating plausible motions for a diverse range of characters is a long‐standing goal in computer graphics. Current learning‐based motion synthesis methods rely on large‐scale motion datasets, which are often difficult if not impossible to acquire. On the other hand, pose data is more accessible, since static posed characters are easier to create and can even be extracted from images using recent advancements in computer vision. In this paper, we tap into this alternative data source and introduce a neural motion synthesis approach through retargeting, which generates plausible motion of various characters that only have pose data by transferring motion from one single existing motion capture dataset of another drastically different characters. Our experiments show that our method effectively combines the motion features of the source character with the pose features of the target character, and performs robustly with small or noisy pose data sets, ranging from a few artist‐created poses to noisy poses estimated directly from images. Additionally, a conducted user study indicated that a majority of participants found our retargeted motion to be more enjoyable to watch, more lifelike in appearance, and exhibiting fewer artifacts. Our code and dataset can be accessed here.

Dual Leap Motion Controller 2: A Robust Dataset for Multi-view Hand Pose Recognition

This paper presents Multi-view Leap2 Hand Pose Dataset (ML2HP Dataset), a new dataset for hand pose recognition, captured using a multi-view recording setup with two Leap Motion Controller 2 devices. This dataset encompasses a diverse range of hand poses, recorded from different angles to ensure comprehensive coverage. The dataset includes real images with the associated precise and automatic hand properties, such as landmark coordinates, velocities, orientations, and finger widths. This dataset has been meticulously designed and curated to maintain a balance in terms of subjects, hand poses, and the usage of right or left hand, ensuring fairness and parity. The content includes 714,000 instances from 21 subjects of 17 different hand poses (including real images and 247 associated hand properties). The multi-view setup is necessary to mitigate hand occlusion phenomena, ensuring continuous tracking and pose estimation required in real human-computer interaction applications. This dataset contributes to advancing the field of multimodal hand pose recognition by providing a valuable resource for developing advanced artificial intelligence human computer interfaces.

Diverse Humanoid Robot Pose Estimation from Images Using Only Sparse Datasets

We present a novel dataset for humanoid robot pose estimation from images, addressing the critical need for accurate pose estimation to enhance human–robot interaction in extended reality (XR) applications. Despite the importance of this task, large-scale pose datasets for diverse humanoid robots remain scarce. To overcome this limitation, we collected sparse pose datasets for commercially available humanoid robots and augmented them through various synthetic data generation techniques, including AI-assisted image synthesis, foreground removal, and 3D character simulations. Our dataset is the first to provide full-body pose annotations for a wide range of humanoid robots exhibiting diverse motions, including side and back movements, in real-world scenarios. Furthermore, we introduce a new benchmark method for real-time full-body 2D keypoint estimation from a single image. Extensive experiments demonstrate that our extended dataset-based pose estimation approach achieves over 33.9% improvement in accuracy compared to using only sparse datasets. Additionally, our method demonstrates the real-time capability of 42 frames per second (FPS) and maintains full-body pose estimation consistency in side and back motions across 11 differently shaped humanoid robots, utilizing approximately 350 training images per robot.

EvHandPose: Event-Based 3D Hand Pose Estimation With Sparse Supervision.

Event camera shows great potential in 3D hand pose estimation, especially addressing the challenges of fast motion and high dynamic range in a low-power way. However, due to the asynchronous differential imaging mechanism, it is challenging to design event representation to encode hand motion information especially when the hands are not moving (causing motion ambiguity), and it is infeasible to fully annotate the temporally dense event stream. In this paper, we propose EvHandPose with novel hand flow representations in Event-to-Pose module for accurate hand pose estimation and alleviating the motion ambiguity issue. To solve the problem under sparse annotation, we design contrast maximization and hand-edge constraints in Pose-to-IWE (Image with Warped Events) module and formulate EvHandPose in a weakly-supervision framework. We further build EvRealHands, the first large-scale real-world event-based hand pose dataset on several challenging scenes to bridge the real-synthetic domain gap. Experiments on EvRealHands demonstrate that EvHandPose outperforms previous event-based methods under all evaluation scenes, achieves accurate and stable hand pose estimation with high temporal resolution in fast motion and strong light scenes compared with RGB-based methods, generalizes well to outdoor scenes and another type of event camera, and shows the potential for the hand gesture recognition task.

A multi-channel spatial information feature based human pose estimation algorithm

Human pose estimation is an important task in computer vision, which can provide key point detection of human body and obtain bone information. At present, human pose estimation is mainly utilized for detection of large targets, and there is no solution for detection of small targets. This paper proposes a multi-channel spatial information feature based human pose (MCSF-Pose) estimation algorithm to address the issue of medium and small targets inaccurate detection of human key points in scenarios involving occlusion and multiple poses. The MCSF-Pose network is a bottom-up regression network. Firstly, an UP-Focus module is designed to expand the feature information while reducing parameter computation during the up-sampling process. Then, the channel segmentation strategy is adopted to cut the features, and the feature information of multiple dimensions is retained through different convolutional groups, which reduces the parameter lightweight network model and makes up for the loss of the feature information associated with the depth of the network. Finally, the three-layer PANet structure is designed to reduce the complexity of the model. With the aid of the structure, it also to improve the detection accuracy and anti-interference ability of human key points. The experimental results indicate that the proposed algorithm outperforms YOLO-Pose and other human pose estimation algorithms on COCO2017 and MPII human pose datasets.

Multi-Task Intelligent Monitoring of Construction Safety Based on Computer Vision

Effective safety management is vital for ensuring construction safety. Traditional safety inspections in construction heavily rely on manual labor, which is both time-consuming and labor-intensive. Extensive research has been conducted integrating computer-vision technologies to facilitate intelligent surveillance and improve safety measures. However, existing research predominantly focuses on singular tasks, while construction environments necessitate comprehensive analysis. This study introduces a multi-task computer vision technology approach for the enhanced monitoring of construction safety. The process begins with the collection and processing of multi-source video surveillance data. Subsequently, YOLOv8, a deep learning-based computer vision model, is adapted to meet specific task requirements by modifying the head component of the framework. This adaptation enables efficient detection and segmentation of construction elements, as well as the estimation of person and machine poses. Moreover, a tracking algorithm integrates these capabilities to continuously monitor detected elements, thereby facilitating the proactive identification of unsafe practices on construction sites. This paper also presents a novel Integrated Excavator Pose (IEP) dataset designed to address the common challenges associated with different single datasets, thereby ensuring accurate detection and robust application in practical scenarios.

Overcoming Data Deficiency for Multi-Person Pose Estimation.

Building multi-person pose estimation (MPPE) models that can handle complex foreground and uncommon scenes is an important challenge in computer vision. Aside from designing novel models, strengthening training data is a promising direction but remains largely unexploited for the MPPE task. In this article, we systematically identify the key deficiencies of existing pose datasets that prevent the power of well-designed models from being fully exploited and propose the corresponding solutions. Specifically, we find that the traditional data augmentation techniques are inadequate in addressing the two key deficiencies, imbalanced instance complexity (IC) (evaluated by our new metric IC) and insufficient realistic scenes. To overcome these deficiencies, we propose a model-agnostic full-view data generation (Full-DG) method to enrich the training data from the perspectives of both poses and scenes. By hallucinating images with more balanced pose complexity and richer real-world scenes, Full-DG can help improve pose estimators' robustness and generalizability. In addition, we introduce a plug-and-play adaptive category-aware loss (AC-loss) to alleviate the severe pixel-level imbalance between keypoints and backgrounds (i.e., around 1:600). Full-DG together with AC-loss can be readily applied to both the bottom-up and top-down models to improve their accuracy. Notably, plugging into the representative estimators HigherHRNet and HRNet, our method achieves substantial performance gains of 1.0%-2.9% AP on the COCO benchmark, and 1.0%-5.1% AP on the CrowdPose benchmark.

Machine Learning-Based Stroke Patient Rehabilitation Stage Classification Using Kinect Data

Everyone aspires to live a healthy life, but many will inevitably experience some form of disease, illness, or accident that results in disability at some point. Rehabilitation plays a crucial role in helping individuals recover from these disabilities and return to their daily activities. Traditional rehabilitation methods are often expensive, are inefficient, and lead to slow progress for patients. However, in this era of technology, various sensor-based automatic rehabilitation is also possible. A Kinect sensor is a skeletal tracking device that captures human motions and gestures. It can provide feedback to the users, allowing them to better understand their progress and adjust their movements accordingly. In this study, stroke-based rehabilitation is presented along with the Toronto Rehab Stroke Pose Dataset (TRSP). Pre-processing of the raw dataset was performed using various features, and several state-of-the-art classifiers were applied to evaluate the data provided by the Kinect sensor. Among the various classifiers, eXtreme Gradient Boosing (XGB) attained the maximum accuracy of 92% for the TRSP dataset. Furthermore, hyperparameters of the XGB have been optimized using a metaheuristic gray wolf optimizer for better performance.

GAN-BodyPose: Real-time 3D human body pose data key point detection and quality assessment assisted by generative adversarial network

With the rapid advancement of deep learning and computer vision, these technologies are becoming increasingly vital in areas like virtual reality, medical diagnosis, and sports training. Existing methods for real-time 3D human body pose keypoint detection and quality assessment face significant challenges such as insufficient detection accuracy, low computational efficiency, and high data quality requirements. To address these challenges, we propose an innovative solution, GAN-BodyPose. This approach integrates 3D convolutional neural networks, self-attention mechanisms, and generative adversarial networks to deliver efficient and accurate detection and assessment in real time. The GAN-BodyPose framework combines 3D-CNN and self-attention for effective feature extraction and keypoint detection, enhanced further by generative adversarial networks for superior data quality and accuracy. Our extensive evaluations using a large-scale 3D human body pose dataset demonstrated that GAN-BodyPose outperforms traditional methods, showing improvements in processing speed (15% faster), accuracy in terms of Mean Per Joint Position Error (reduced by approximately 2.2%), and an Area Under the Curve (AUC) score increased by approximately 9.5% compared to HR-Net and other datasets. Additionally, it achieves lower Floating-Point Operations (FLOPs) by about 9.3%, indicating more efficient computational performance. These advancements underline the potential of our approach to significantly enhance user experiences in virtual reality, motion capture, and other real-time applications. The successful application of GAN-BodyPose promises greater efficiency and precision in fields ranging from game development to medical diagnostics, and robust support for human-computer interaction and gesture recognition. This research represents a substantial contribution to deep learning applications in robot control, decision-making, and broadens the research foundation in these domains.

Data-driven ergonomic assessment of construction workers

Automating ergonomic assessment improves both objectivity and cost-effectiveness. However, existing ergonomic scales remain susceptible to inaccuracy and insensitivity when assessing diversified construction activities. This susceptibility stems from the unreliable ranges, sharp boundaries, and oversimplified binary rules within conventional joint-level assessment rules, coupled with the restricted transformation rules to integer input/output. This paper presents a data-driven ergonomic assessment method grounded in statistical data from Construction Motion Data Library (CML) dataset, comprising: 1) Developing joint-level scoring models to primarily improve accuracy through leveraging Heuristics Gaussian Cloud Transformation (H-GCT), and 2) Designing fuzzy inference mechanism to enhance sensitivity through retaining risk information during transformation across levels. Subsequently, a three-step validation confirms accurate risk identification, sensitive risk feedback, and positive correlation with previous methods, contributing to occupational health and safety management. Future research could enhance the method by expanding the pose dataset, recruiting more participants, and refining the 3D pose estimation.

Test Platform for Developing New Optical Position Tracking Technology towards Improved Head Motion Correction in Magnetic Resonance Imaging.

Optical tracking of head pose via fiducial markers has been proven to enable effective correction of motion artifacts in the brain during magnetic resonance imaging but remains difficult to implement in the clinic due to lengthy calibration and set up times. Advances in deep learning for markerless head pose estimation have yet to be applied to this problem because of the sub-millimetre spatial resolution required for motion correction. In the present work, two optical tracking systems are described for the development and training of a neural network: one marker-based system (a testing platform for measuring ground truth head pose) with high tracking fidelity to act as the training labels, and one markerless deep-learning-based system using images of the markerless head as input to the network. The markerless system has the potential to overcome issues of marker occlusion, insufficient rigid attachment of the marker, lengthy calibration times, and unequal performance across degrees of freedom (DOF), all of which hamper the adoption of marker-based solutions in the clinic. Detail is provided on the development of a custom moiré-enhanced fiducial marker for use as ground truth and on the calibration procedure for both optical tracking systems. Additionally, the development of a synthetic head pose dataset is described for the proof of concept and initial pre-training of a simple convolutional neural network. Results indicate that the ground truth system has been sufficiently calibrated and can track head pose with an error of <1 mm and <1°. Tracking data of a healthy, adult participant are shown. Pre-training results show that the average root-mean-squared error across the 6 DOF is 0.13 and 0.36 (mm or degrees) on a head model included and excluded from the training dataset, respectively. Overall, this work indicates excellent feasibility of the deep-learning-based approach and will enable future work in training and testing on a real dataset in the MRI environment.

Animal Pose Estimation Based on Contrastive Learning with Dynamic Conditional Prompts.

Traditional animal pose estimation techniques based on images face significant hurdles, including scarce training data, costly data annotation, and challenges posed by non-rigid deformation. Addressing these issues, we proposed dynamic conditional prompts for the prior knowledge of animal poses in language modalities. Then, we utilized a multimodal (language-image) collaborative training and contrastive learning model to estimate animal poses. Our method leverages text prompt templates and image feature conditional tokens to construct dynamic conditional prompts that integrate rich linguistic prior knowledge in depth. The text prompts highlight key points and relevant descriptions of animal poses, enhancing their representation in the learning process. Meanwhile, transformed via a fully connected non-linear network, image feature conditional tokens efficiently embed the image features into these prompts. The resultant context vector, derived from the fusion of the text prompt template and the image feature conditional token, generates a dynamic conditional prompt for each input sample. By utilizing a contrastive language-image pre-training model, our approach effectively synchronizes and strengthens the training interactions between image and text features, resulting in an improvement to the precision of key-point localization and overall animal pose estimation accuracy. The experimental results show that language-image contrastive learning based on dynamic conditional prompts enhances the average accuracy of animal pose estimation on the AP-10K and Animal Pose datasets.

Virtual Analysis for Spinal Cord Injury Rehabilitation

Background Spinal cord injuries (SCI) are debilitating conditions affecting individuals worldwide annually, leading to physical, emotional, and cognitive challenges. Effective rehabilitation for SCI patients is crucial for restoring motor function and enhancing their overall quality of life. Advances in technology, including machine learning (ML) and computer vision, offer promising avenues for personalized SCI treatment. Aims This paper aimed to propose an automated and cost-effective system for spinal cord injury (SCI) rehabilitation using machine learning techniques, leveraging data from the Toronto Rehab Pose dataset and Mediapipe for real-time tracking. Objective The objective is to develop a system that predicts rehabilitation outcomes for upper body movements, highlighting the transformative role of ML in personalized SCI treatment and offering tailored strategies for improved outcomes. Methods The proposed system utilized data from the Toronto Rehab Pose dataset and Mediapipe for real-time tracking. Machine learning models, including Support Vector Machines (SVM), Logistic Regression, Naive Bayes, and XGBoost, were employed for outcome prediction. Features such as joint positions, angles, velocities, and accelerations were extracted from movement data to train the models. Results Statistical analysis revealed the ability of the system to accurately classify rehabilitation outcomes, with an average accuracy of 98.5%. XGBoost emerged as the top-performing algorithm, demonstrating superior accuracy and precision scores across all exercises. Conclusion This paper emphasizes the importance of continuous monitoring and adjustment of rehabilitation plans based on real-time progress data, highlighting the dynamic nature of SCI rehabilitation and the need for adaptive treatment strategies. By predicting rehabilitation outcomes with high accuracy, the system enables clinicians to devise targeted interventions, optimizing the efficacy of the rehabilitation process.

Seeing through the Tactile

Humans spend about one-third of their lives resting. Reconstructing human dynamics in in-bed scenarios is of considerable significance in sleep studies, bedsore monitoring, and biomedical factor extractions. However, the mainstream human pose and shape estimation methods mainly focus on visual cues, facing serious issues in non-line-of-sight environments. Since in-bed scenarios contain complicated human-environment contact, pressure-sensing bedsheets provide a non-invasive and privacy-preserving approach to capture the pressure distribution on the contact surface, and have shown prospects in many downstream tasks. However, few studies focus on in-bed human mesh recovery. To explore the potential of reconstructing human meshes from the sensed pressure distribution, we first build a high-quality temporal human in-bed pose dataset, TIP, with 152K multi-modality synchronized images. We then propose a label generation pipeline for in-bed scenarios to generate reliable 3D mesh labels with a SMPLify-based optimizer. Finally, we present PIMesh, a simple yet effective temporal human shape estimator to directly generate human meshes from pressure image sequences. We conduct various experiments to evaluate PIMesh's performance, showing that PIMesh archives 79.17mm joint position errors on our TIP dataset. The results demonstrate that the pressure-sensing bedsheet could be a promising alternative for long-term in-bed human shape estimation.

Sign Language Word Detection Using LRCN

Sign language is the most effective communication for deaf or hard-of-hearing people. Specialized training is required to understand sign language, and as such, people without disabilities around them cannot communicate effectively. The main objective of this study is to develop a mechanism for streamlining the deep learning model for sign language recognition by utilizing the 30 most prevalent words in our everyday lives. The dataset was designed through 30 ASL (American Sign Language) words consisting of custom-processed video sequences, which consist of 5 subjects and 50 sample videos for each class. The CNN model can be applied to video frames to extract spatial properties. Using CNN’s acquired data, the LSTM model may then predict the action being performed in the video. We present and evaluate the results of two separate datasets—the Pose dataset and the Raw video dataset. The dataset was trained with the Long-term Recurrent Convolutional Network (LRCN) approach. Finally, a test accuracy of 92.66% was reached for the raw dataset, while 93.66% for the pose dataset.

Hand LightWeightNet: an optimized hand pose estimation for interactive mobile interfaces

In this paper, a hand pose estimation method is introduced that combines MobileNetV3 and CrossInfoNet into a single pipeline. The proposed approach is tailored for mobile phone processors through optimizations, modifications, and enhancements made to both architectures, resulting in a lightweight solution. MobileNetV3 provides the bottleneck for feature extraction and refinements while CrossInfoNet benefits the proposed system through a multitask information sharing mechanism. In the feature extraction stage, we utilized an inverted residual block that achieves a balance between accuracy and efficiency in limited parameters. Additionally, in the feature refinement stage, we incorporated a new best-performing activation function called “activate or not” ACON, which demonstrated stability and superior performance in learning linearly and non-linearly gates of the whole activation area of the network by setting hyperparameters to switch between active and inactive states. As a result, our network operated with 65% reduced parameters, but improved speed by 39% which is suitable for running in a mobile device processor. During experiment, we conducted test evaluation on three hand pose datasets to assess the generalization capacity of our system. On all the tested datasets, the proposed approach demonstrates consistently higher performance while using significantly fewer parameters than existing methods. This indicates that the proposed system has the potential to enable new hand pose estimation applications such as virtual reality, augmented reality and sign language recognition on mobile devices.

Detect Any Keypoints: An Efficient Light-Weight Few-Shot Keypoint Detector

Recently the prompt-based models have become popular across various language and vision tasks. Following that trend, we perform few-shot keypoint detection (FSKD) by detecting any keypoints in a query image, given the prompts formed by support images and keypoints. FSKD can be applied to detecting keypoints and poses of diverse animal species. In order to maintain flexibility of detecting varying number of keypoints, existing FSKD approaches modulate query feature map per support keypoint, then detect the corresponding keypoint from each modulated feature via a detection head. Such a separation of modulation-detection makes model heavy and slow when the number of keypoints increases. To overcome this issue, we design a novel light-weight detector which combines modulation and detection into one step, with the goal of reducing the computational cost without the drop of performance. Moreover, to bridge the large domain shift of keypoints between seen and unseen species, we further improve our model with mean feature based contrastive learning to align keypoint distributions, resulting in better keypoint representations for FSKD. Compared to the state of the art, our light-weight detector reduces the number of parameters by 50%, training/test time by 50%, and achieves 5.62% accuracy gain on 1-shot novel keypoint detection in the Animal pose dataset. Our model is also robust to the number of keypoints and saves memory when evaluating a large number of keypoints (e.g., 1000) per episode.