Multi-person 3D Pose Estimation Research Articles

Multi-person 3D pose estimation from unlabelled data

Its numerous applications make multi-human 3D pose estimation a remarkably impactful area of research. Nevertheless, it presents several challenges, especially when approached using multiple views and regular RGB cameras as the only input. First, each person must be uniquely identified in the different views. Secondly, it must be robust to noise, partial occlusions, and views where a person may not be detected. Thirdly, many pose estimation approaches rely on environment-specific annotated datasets that are frequently prohibitively expensive and/or require specialised hardware. Specifically, this is the first multi-camera, multi-person data-driven approach that does not require an annotated dataset. In this work, we address these three challenges with the help of self-supervised learning. In particular, we present a three-staged pipeline and a rigorous evaluation providing evidence that our approach performs faster than other state-of-the-art algorithms, with comparable accuracy, and most importantly, does not require annotated datasets. The pipeline is composed of a 2D skeleton detection step, followed by a Graph Neural Network to estimate cross-view correspondences of the people in the scenario, and a Multi-Layer Perceptron that transforms the 2D information into 3D pose estimations. Our proposal comprises the last two steps, and it is compatible with any 2D skeleton detector as input. These two models are trained in a self-supervised manner, thus avoiding the need for datasets annotated with 3D ground-truth poses.

YOLO-Rlepose: Improved YOLO Based on Swin Transformer and Rle-Oks Loss for Multi-Person Pose Estimation

In recent years, there has been significant progress in human pose estimation, fueled by the widespread adoption of deep convolutional neural networks. However, despite these advancements, multi-person 2D pose estimation still remains highly challenging due to factors such as occlusion, noise, and non-rigid body movements. Currently, most multi-person pose estimation approaches handle joint localization and association separately. This study proposes a direct regression-based method to estimate the 2D human pose from a single image. The authors name this network YOLO-Rlepose. Compared to traditional methods, YOLO-Rlepose leverages Transformer models to better capture global dependencies between image feature blocks and preserves sufficient spatial information for keypoint detection through a multi-head self-attention mechanism. To further improve the accuracy of the YOLO-Rlepose model, this paper proposes the following enhancements. Firstly, this study introduces the C3 Module with Swin Transformer (C3STR). This module builds upon the C3 module in You Only Look Once (YOLO) by incorporating a Swin Transformer branch, enhancing the YOLO-Rlepose model’s ability to capture global information and rich contextual information. Next, a novel loss function named Rle-Oks loss is proposed. The loss function facilitates the training process by learning the distributional changes through Residual Log-likelihood Estimation. To assign different weights based on the importance of different keypoints in the human body, this study introduces a weight coefficient into the loss function. The experiments proved the efficiency of the proposed YOLO-Rlepose model. On the COCO dataset, the model outperforms the previous SOTA method by 2.11% in AP.

RPM 2.0: RF-Based Pose Machines for Multi-Person 3D Pose Estimation

Advanced human sensing technologies based on radio frequency (RF) signals have gained widespread attention in recent years. However, due to the sparsity and incompleteness of RF signals, fine-grained RF-based multi-person 3D pose estimation has progressed more slowly. In this paper, we present RF-based Pose Machine (RPM 2.0) for multi-person 3D pose estimation using RF signals. Specifically, we first develop a lightweight anchor-free detector module to locate and crop regions of interest from horizontal and vertical RF signals. Afterward, we treat the horizontal and vertical millimeter-wave radars as “RF cameras” with different viewing angles and propose a Multi-view Fusion Network to unproject the RF signals into a unified latent feature space, and then calculate the correlation for weighted fusion. Finally, a Spatio-Temporal Attention Network is designed to reconstruct the multi-person 3D skeleton sequences, in which the spatial attention module is proposed to recover invisible body parts using non-local correlations among joints and the temporal attention module refines the 3D pose sequences using temporal coherency learned from frame queries. We evaluate the performance of the proposed RPM 2.0 and state-of-the-art methods on a large-scale dataset with multi-person 3D pose labels and corresponding radar signals. The experimental results show that RPM 2.0 outperforms all of the baseline methods, which locates multi-person 3D key points with an average error of 73 <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">mm</i> and generalizes well in new data such as occlusion, low illumination.

Multi-Person 3D Pose Esitmation With Occlusion Reasoning

The performance of existing methods for multi-person 3D pose estimation in crowded scenes is still limited, due to the challenge of heavy overlapping among persons. Attempt to address this issue, we propose a progressive inference scheme, i.e., Articulation-aware Knowledge Exploration (AKE), to improve the multi-person 3D pose models on those samples with complex occlusions at the inference stage. We argue it is beneficial to explore the underlying articulated information/knowledge of the human body, which helps to further correct the predicted poses in those samples. To exploit such information, we propose an iterative scheme to achieve a self-improving loop for keypoint association. Specifically, we introduce a kinematic validation module for locating unreasonable articulations and an occluded-keypoint discovering module for discovering occluded articulations. Extensive experiments on two challenging benchmarks under both weakly-supervised and fully-supervised settings demonstrate the superiority and generalization ability of our proposed method for crowded scenes.

Enhanced 3D Pose Estimation in Multi-Person, Multi-View Scenarios through Unsupervised Domain Adaptation with Dropout Discriminator.

Data-driven pose estimation methods often assume equal distributions between training and test data. However, in reality, this assumption does not always hold true, leading to significant performance degradation due to distribution mismatches. In this study, our objective is to enhance the cross-domain robustness of multi-view, multi-person 3D pose estimation. We tackle the domain shift challenge through three key approaches: (1) A domain adaptation component is introduced to improve estimation accuracy for specific target domains. (2) By incorporating a dropout mechanism, we train a more reliable model tailored to the target domain. (3) Transferable Parameter Learning is employed to retain crucial parameters for learning domain-invariant data. The foundation for these approaches lies in the H-divergence theory and the lottery ticket hypothesis, which are realized through adversarial training by learning domain classifiers. Our proposed methodology is evaluated using three datasets: Panoptic, Shelf, and Campus, allowing us to assess its efficacy in addressing domain shifts in multi-view, multi-person pose estimation. Both qualitative and quantitative experiments demonstrate that our algorithm performs well in two different domain shift scenarios.

Snipper: A Spatiotemporal Transformer for Simultaneous Multi-Person 3D Pose Estimation Tracking and Forecasting on a Video Snippet

Multi-person pose understanding from RGB videos involves three complex tasks: pose estimation, tracking and motion forecasting. Intuitively, accurate multi-person pose estimation facilitates robust tracking, and robust tracking builds crucial history for correct motion forecasting. Most existing works either focus on a single task or employ multi-stage approaches to solving multiple tasks separately, which tends to make sub-optimal decision at each stage and also fail to exploit correlations among the three tasks. In this paper, we propose Snipper, a unified framework to perform multi-person 3D pose estimation, tracking, and motion forecasting simultaneously in a single stage. We propose an efficient yet powerful deformable attention mechanism to aggregate spatiotemporal information from the video snippet. Building upon this deformable attention, a video transformer is learned to encode the spatiotemporal features from the multi-frame snippet and to decode informative pose features for multi-person pose queries. Finally, these pose queries are regressed to predict multi-person pose trajectories and future motions in a single shot. In the experiments, we show the effectiveness of Snipper on three challenging public datasets where our generic model rivals specialized state-of-art baselines for pose estimation, tracking, and forecasting. Code is available at https://github.com/JimmyZou/Snipper.

A perception-enhancement network for accurate multi-person 2D pose estimation

A perception-enhancement network for accurate multi-person 2D pose estimation

Ergonomic assessment based on monocular RGB camera in elderly care by a new multi-person 3D pose estimation technique (ROMP)

Nursing staff members are at high risk of work-related musculoskeletal disorders (WMSDs), which not only threaten their health but also impact the quality of elderly care. Ergonomic posture risk assessment (EPRA) is usually employed to identify potential WMSD risks such as extreme posture and repetitive movements. A monocular RGB camera has been used for the EPRA in recent years due to its short time requirements and low cost. However, most work scenarios do not involve multi-person situations.Therefore, based on the latest 3D pose estimation algorithm—Monocular, One-stage Regression of Multiple 3D People (ROMP)—this study proposes a method that uses one monocular RGB camera to conduct the EPRA in multi-person and occluded scenarios. The accuracy of our method was calculated through 12 care tasks involving multi-person and occlusion, using the Noitom motion capture (MoCap) system.The results show that our method performed well, with an average accuracy of 83.8% and 90.7%, respectively, using two EPRA scoring tools, RULA and OWAS. The mean absolute error (MAE) of each joint angle was 9.4°. Thus, ROMP seems to be a potential method for conducting the EPRA in nursing workspaces with unsatisfactory conditions using a single monocular RGB camera.

Bottom-up 2D pose estimation via dual anatomical centers for small-scale persons

In multi-person 2D pose estimation, the bottom-up methods simultaneously predict poses for all persons, and unlike the top-down methods, do not rely on human detection. However, the SOTA bottom-up methods’ accuracy is still inferior compared to the existing top-down methods. This is due to the predicted human poses being regressed based on the inconsistent human bounding box center and the lack of human-scale normalization, leading to the predicted human poses being inaccurate and small-scale persons being missed. To push the envelope of the bottom-up pose estimation, we firstly propose multi-scale training to enhance the network to handle scale variation with single-scale testing, particularly for small-scale persons. Secondly, we introduce dual anatomical centers (i.e., head and body), where we can predict the human poses more accurately and reliably, especially for small-scale persons. Moreover, existing bottom-up methods use multi-scale testing to boost the accuracy of pose estimation at the price of multiple additional forward passes, which weakens the efficiency of bottom-up methods, the core strength compared to top-down methods. By contrast, our multi-scale training enables the model to predict high-quality poses in a single forward pass (i.e., single-scale testing). Our method achieves 38.4% improvement on bounding box precision and 39.1% improvement on bounding box recall over the state of the art (SOTA) on the challenging small-scale persons subset of COCO. For the human pose AP evaluation, we achieve a new SOTA (71.0 AP) on the COCO test-dev set with the single-scale testing. We also achieve the top performance (40.3 AP) on the OCHuman dataset in cross-dataset evaluation.

Dual Networks Based 3D Multi-Person Pose Estimation From Monocular Video.

Monocular 3D human pose estimation has made progress in recent years. Most of the methods focus on single persons, which estimate the poses in the person-centric coordinates, i.e., the coordinates based on the center of the target person. Hence, these methods are inapplicable for multi-person 3D pose estimation, where the absolute coordinates (e.g., the camera coordinates) are required. Moreover, multi-person pose estimation is more challenging than single pose estimation, due to inter-person occlusion and close human interactions. Existing top-down multi-person methods rely on human detection (i.e., top-down approach), and thus suffer from the detection errors and cannot produce reliable pose estimation in multi-person scenes. Meanwhile, existing bottom-up methods that do not use human detection are not affected by detection errors, but since they process all persons in a scene at once, they are prone to errors, particularly for persons in small scales. To address all these challenges, we propose the integration of top-down and bottom-up approaches to exploit their strengths. Our top-down network estimates human joints from all persons instead of one in an image patch, making it robust to possible erroneous bounding boxes. Our bottom-up network incorporates human-detection based normalized heatmaps, allowing the network to be more robust in handling scale variations. Finally, the estimated 3D poses from the top-down and bottom-up networks are fed into our integration network for final 3D poses. To address the common gaps between training and testing data, we do optimization during the test time, by refining the estimated 3D human poses using high-order temporal constraint, re-projection loss, and bone length regularizations. We also introduce a two-person pose discriminator that enforces natural two-person interactions. Finally, we apply a semi-supervised method to overcome the 3D ground-truth data scarcity. Our evaluations demonstrate the effectiveness of the proposed method and its individual components. Our code and pretrained models are available publicly: https://github.com/3dpose/3D-Multi-Person-Pose.

VoxelTrack: Multi-Person 3D Human Pose Estimation and Tracking in the Wild.

We present VoxelTrack for multi-person 3D pose estimation and tracking from a few cameras which are separated by wide baselines. It employs a multi-branch network to jointly estimate 3D poses and re-identification (Re-ID) features for all people in the environment. In contrast to previous efforts which require to establish cross-view correspondence based on noisy 2D pose estimates, it directly estimates and tracks 3D poses from a 3D voxel-based representation constructed from multi-view images. We first discretize the 3D space by regular voxels and compute a feature vector for each voxel by averaging the body joint heatmaps that are inversely projected from all views. We estimate 3D poses from the voxel representation by predicting whether each voxel contains a particular body joint. Similarly, a Re-ID feature is computed for each voxel which is used to track the estimated 3D poses over time. The main advantage of the approach is that it avoids making any hard decisions based on individual images. The approach can robustly estimate and track 3D poses even when people are severely occluded in some cameras. It outperforms the state-of-the-art methods by a large margin on four public datasets including Shelf, Campus, Human3.6 M and CMU Panoptic.

Fast and Robust Multi-Person 3D Pose Estimation and Tracking From Multiple Views.

This paper addresses the problem of reconstructing 3D poses of multiple people from a few calibrated camera views. The main challenge of this problem is to find the cross-view correspondences among noisy and incomplete 2D pose predictions. Most previous methods address this challenge by directly reasoning in 3D using a pictorial structure model, which is inefficient due to the huge state space. We propose a fast and robust approach to solve this problem. Our key idea is to use a multi-way matching algorithm to cluster the detected 2D poses in all views. Each resulting cluster encodes 2D poses of the same person across different views and consistent correspondences across the keypoints, from which the 3D pose of each person can be effectively inferred. The proposed convex optimization based multi-way matching algorithm is efficient and robust against missing and false detections, without knowing the number of people in the scene. Moreover, we propose to combine geometric and appearance cues for cross-view matching. Finally, an efficient tracking method is proposed to track the detected 3D poses across the multi-view video. The proposed approach achieves the state-of-the-art performance on the Campus and Shelf datasets, while being efficient for real-time applications.

Multi-person 3D pose estimation from a single image captured by a fisheye camera

Multi-person 3D pose estimation with absolute depths for a fisheye camera is a challenging task but with valuable applications in daily life, especially for video surveillance. However, to the best of our knowledge, such problem has not been explored so far, leaving a gap in practical applications. In this work, we first propose a method for multi-person 3D pose estimation from a single image taken by a fisheye camera. Our method consists of two branches to estimate absolute 3D human poses: (1) a 2D-to-3D lifting module to predict root-relative 3D human poses (HPoseNet); (2) a root regression module to estimate absolute root locations in the camera coordinate (HRootNet). Finally, we propose a fisheye re-projection module without using ground-truth camera parameters to connect two branches, alleviating the impact of image distortions on 3D pose estimation and further regularizing prediction absolute 3D poses. Experimental results demonstrate that our method achieves the state-of-the-art performance on two public multi-person 3D pose datasets with synthetic fisheye images and our newly collected dataset with real fisheye images. The code and new dataset will be made publicly available.

Multi-task neural network with physical constraint for real-time multi-person 3D pose estimation from monocular camera

3D human pose estimation has many important applications in human-computer interaction and human action recognition. Simultaneously achieving real-time speed, varying human number, and high accuracy from a single RGB image is a challenging problem. To this end, this paper proposes a multi-task and multi-level neural network structure with physical constraint. The unique network structure estimates 3D human poses from single RGB image in an end-to-end way and achieves both high accuracy and high speed. Experimental results shows that the proposed system achieves 21 fps on RTX 2080 GPU with only 33 mm accuracy loss compared with conventional works. The mechanism of the network is also analyzed through network visualization. This work shows the possibility of estimating 3D human pose from a single RGB monocular camera with real-time speed.

Double anchor embedding for accurate multi-person 2D pose estimation

Multi-person pose estimation is an important field in computer vision. Due to the lower time complexity, the bottom-up approaches have recently received more attention in multi-person 2D pose estimation, however, they are more sensitive to challenges in real-world scenarios. In this paper, we propose a multi-person pose estimation algorithm based on the Double Anchor Embedding (DAE), which shows that bottom-up algorithms are still competitive in precision. Firstly, for reducing the modeling difficulty of the detection task we divide the human joints into upper and lower half groups which are internally continuous and highly correlated. Accordingly, a novel joint affinity cue, called Double Anchor Embedding is designed, which can help the network effectively extract the information of both local contexts and global contexts, so that can better cope with occluded scenes and complex postures. Secondly, the parallel greedy joint inference algorithm is proposed to alleviate the mismatching problem of distant joints in the post-processing stage, which can also accelerate the matching process to some extent. Extensive experiments on two challenging datasets demonstrate the effectiveness and potential of our proposed framework, which is comparable to the current state-of-the-art methods.

AnimePose: Multi-person 3D pose estimation and animation

3D animation of human body movement is quite challenging as it involves using a huge setup with several motion trackers all over the persons body to track the movements of every limb. This is time- consuming and may cause the person discomfort in wearing exoskeleton body suits with motion sensors. In this work, we present a trivial yet effective solution to generate simple 3D animation of human movement of multiple persons from a 2D video using deep learning. Although significant improvement has been achieved recently in 3D human pose estimation, most of the prior works work well in case of single person pose estimation and multi-person pose estimation is still a challenging problem. In this work, we firstly propose a supervised multi-person 3D pose estimation and animation framework namely AnimePose for a given input RGB video sequence. The pipeline of the proposed system consists of various modules: i) Multi-Person 2D pose estimation, ii) Depth Map estimation, iii) Lifting 2D poses to 3D poses, iv) Person trajectory prediction and human pose tracking. Our proposed system produces comparable results on previous state-of-the-art 3D multi-person pose estimation methods on publicly available dataset MuPoTS-3D dataset and it also outperforms previous competing human pose tracking methods by a significant margin of 11.7% performance gain on MOTA score on Posetrack 2018 dataset.

Robustify Hand Tracking by Fusing Generative and Discriminative Methods

With the development of virtual reality (VR) technology and its applications in many fields, creating simulated hands in the virtual environment is an e ective way to replace the controller as well as to enhance user experience in interactive processes. Therefore, hand tracking problem is gaining a lot of research attention, making an important contribution in recognizing hand postures as well as tracking hand motions for VR’s input or human machine interaction applications. In order to create a markerless real-time hand tracking system suitable for natural human machine interaction, we propose a new method that combines generative and discriminative methods to solve the hand tracking problem using a single RGBD camera. Our system removes the requirement of the user having to wear to color wrist band and robustifies the hand localization even in di cult tracking scenarios.&#x0D; KeywordsHand tracking, generative method, discriminative method, human performance capture&#x0D; References[1] Malik, A. Elhayek, F. Nunnari, K. Varanasi, Tamaddon, A. Heloir, D. Stricker, Deephps: End-to-end estimation of 3d hand pose and shape by learning from synthetic depth, CoRR abs/1808.09208, 2018.&#x0D; URL http://arxiv.org/abs/1808.09208.&#x0D; [2] Glauser, S. Wu, D. Panozzo, O. Hilliges, Sorkine-Hornung, Interactive hand pose estimation using a stretch-sensing soft glove, ACM Trans, Graph. 38(4) (2019) 1-15.[3] Jiang, H. Xia, C. Guo, A model-based system for real-time articulated hand tracking using a simple data glove and a depth camera, Sensors 19 (2019) 4680. https://doi.org/10.3390/s19214680.[4] Cao, G. Hidalgo, T. Simon, S. Wei, Y. Sheikh, Openpose: Realtime multi-person 2d pose estimation using part a nity fields, CoRR abs/1812.08008, 2018.[5] Tagliasacchi, M. Schroder, A. Tkach, S. Bouaziz, M. Botsch, M. Pauly, Robust articulated-icp for real-time hand tracking, Computer Graphics Forum 34, 2015.[6] Qian, X. Sun, Y. Wei, X. Tang, J. Sun, Realtime and robust hand tracking from depth, in: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014.[7] Tomasi, Petrov, Sastry, 3d tracking = classification + interpolation, in: Proceedings Ninth IEEE International Conference on Computer Vision 2 (2003) 1441-1448.[8] Sharp, C. Keskin, D. Robertson, J. Taylor, J. Shotton, D. Kim, C. Rhemann, I. Leichter, A. Vinnikov, Y. Wei, et al., Accurate, robust, and flexible real-time hand tracking, in: Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems, 2015, pp. 3633-3642.[9] Sridhar, F. Mueller, A. Oulasvirta, C. Theobalt, Fast and robust hand tracking using detection-guided optimization, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015.[10] Oikonomidis, N. Kyriazis, A.A. Argyros, Tracking the articulated motion of two strongly interacting hands, in: 2012 IEEE Conference on Computer Vision and Pattern Recognition, 2012, pp. 1862-1869.[11] Melax, L. Keselman, S. Orsten, Dynamics based 3d skeletal hand tracking, CoRR abs/1705.07640, 2017.[12] Wang, S. Paris, J. Popovic, 6d hands: Markerless hand tracking for computer aided design, 2011, pp. 549-558.&#x0D; https://doi.org/10.1145/2047196.2047269.[13]Tang, T. Yu, T. Kim, Real-time articulated hand pose estimation using semi-supervised transductive regression forests, in: 2013 IEEE International Conference on Computer Vision, 2013, pp. 3224-3231.[14] Oberweger, P. Wohlhart, V. Lepetit, Generalized feedback loop for joint hand-object pose estimation, 2019, CoRR abs/1903.10883. URL http://arxiv.org/abs/1903.10883.[15] Malik, A. Elhayek, F. Nunnari, K. Varanasi, K. Tamaddon, A. Heloir,´ D. Stricker, Deephps: End-to-end estimation of 3d hand pose and shape by learning from synthetic depth, 2018, pp. 110-119. https://doi.org/10.1109/3DV.2018.00023.[16] A. Mohammed, J.L.M. Islam, A deep learning-based end-to-end composite system for hand detection and gesture recognition, Sensors 19 (2019) 5282. https://doi.org/10.3390/s19235282.

A Framework for Fall Detection Based on OpenPose Skeleton and LSTM/GRU Models

Falling is one of the causes of accidental death of elderly people over 65 years old in Taiwan. If the fall incidents are not detected in a timely manner, it could lead to serious injury or even death of those who fell. General fall detection approaches require the users to wear sensors, which could be cumbersome for the users to put on, and misalignment of sensors could lead to erroneous readings. In this paper, we propose using computer vision and applied machine-learning algorithms to detect fall without any sensors. We applied OpenPose real-time multi-person 2D pose estimation to detect movement of a subject using two datasets of 570 × 30 frames recorded in five different rooms from eight different viewing angles. The system retrieves the locations of 25 joint points of the human body and detects human movement through detecting the joint point location changes. The system is able to effectively identify the joints of the human body as well as filtering ambient environmental noise for an improved accuracy. The use of joint points instead of images improves the training time effectively as well as eliminating the effects of traditional image-based approaches such as blurriness, light, and shadows. This paper uses single-view images to reduce equipment costs. We experimented with time series recurrent neural network, long- and short-term memory, and gated recurrent unit models to learn the changes in human joint points in continuous time. The experimental results show that the fall detection accuracy of the proposed model is 98.2%, which outperforms the baseline 88.9% with 9.3% improvement.

OpenPose: Realtime Multi-Person 2D Pose Estimation Using Part Affinity Fields.

Realtime multi-person 2D pose estimation is a key component in enabling machines to have an understanding of people in images and videos. In this work, we present a realtime approach to detect the 2D pose of multiple people in an image. The proposed method uses a nonparametric representation, which we refer to as Part Affinity Fields (PAFs), to learn to associate body parts with individuals in the image. This bottom-up system achieves high accuracy and realtime performance, regardless of the number of people in the image. In previous work, PAFs and body part location estimation were refined simultaneously across training stages. We demonstrate that a PAF-only refinement rather than both PAF and body part location refinement results in a substantial increase in both runtime performance and accuracy. We also present the first combined body and foot keypoint detector, based on an internal annotated foot dataset that we have publicly released. We show that the combined detector not only reduces the inference time compared to running them sequentially, but also maintains the accuracy of each component individually. This work has culminated in the release of OpenPose, the first open-source realtime system for multi-person 2D pose detection, including body, foot, hand, and facial keypoints.

Efficient Convolutional Neural Networks for Depth-Based Multi-Person Pose Estimation

Achieving robust multi-person 2D body landmark localization and pose estimation is essential for human behavior and interaction understanding as encountered for instance in HRI settings. Accurate methods have been proposed recently, but they usually rely on rather deep Convolutional Neural Network (CNN) architecture, thus requiring large computational and training resources. In this paper, we investigate different architectures and methodologies to address these issues and achieve fast and accurate multi-person 2D pose estimation. To foster speed, we propose to work with depth images, whose structure contains sufficient information about body landmarks while being simpler than textured color images and thus potentially requiring less complex CNNs for processing. In this context, we make the following contributions. i) we study several CNN architecture designs combining pose machines relying on the cascade of detectors concept with lightweight and efficient CNN structures; ii) to address the need for large training datasets with high variability, we rely on semi-synthetic data combining multi-person synthetic depth data with real sensor backgrounds; iii) we explore domain adaptation techniques to address the performance gap introduced by testing on real depth images; iv) to increase the accuracy of our fast lightweight CNN models, we investigate knowledge distillation at several architecture levels which effectively enhance performance. Experiments and results on synthetic and real data highlight the impact of our design choices, providing insights into methods addressing standard issues normally faced in practical applications, and resulting in architectures effectively matching our goal in both performance and speed.