Assessment of the Accuracy of a Deep Learning Algorithm- and Video-based Motion Capture System in Estimating Snatch Kinematics Authors: FEDERICO THIELE, FLORIAN PATERNOSTER, CHRIS HUMMEL, FABIAN STÖCKER, DENIS HOLZER Published: 2024 ABSTRACT In weightlifting, quantitative kinematic analysis is essential for evaluating snatch performance. While marker-based (MB) approaches are commonly used, they are impractical for training or competitions. Markerless video-based (VB) systems utilizing deep learning-based pose estimation algorithms could address this issue. This study assessed the comparability and applicability of VB systems in obtaining snatch kinematics by comparing the outcomes to an MB reference system. 21 weightlifters (15 Male, 6 Female) performed 2–3 snatches at 65%, 75%, and 80% of their one-repetition maximum. Snatch kinematics were analyzed using an MB (Vicon Nexus) and VB (Contemplas along with Theia3D) system. Analysis of 131 trials revealed that corresponding lower limb joint center positions of the systems on average differed by 4.7 ± 1.2 cm, and upper limb joint centers by 5.7 ± 1.5 cm. VB and MB lower limb joint angles showed highest agreement in the frontal plane (root mean square difference (RMSD): 11.2 ± 5.9°), followed by the sagittal plane (RMSD: 13.6 ± 4.7°). Statistical Parametric Mapping analysis revealed significant differences throughout most of the movement for all degrees of freedom. Maximum extension angles and velocities during the second pull displayed significant differences (p < .05) for the lower limbs. Our data showed significant differences in estimated kinematics between both systems, indicating a lack of comparability. These differences are likely due to differing models and assumptions, rather than measurement accuracy. However, given the rapid advancements of neural network-based approaches, it holds promise to become a suitable alternative to MB systems in weightlifting analysis. INTRODUCTION The snatch is a discipline in weightlifting, requiring the athlete to lift a barbell off the ground to a straight-arm overhead position in one continuous movement and stand in a fully extended and upright position. Thereby, quantitative motion analysis serves as an effective tool for evaluating and improving lifting techniques (31). In particular, joint angles, joint angular displacements and velocities, and angle amplitudes are of interest throughout the lift in weightlifting analysis (5). For this, marker-based three-dimensional (3D) motion capture systems (MB) using infrared cameras are mostly used to investigate the snatch under laboratory conditions (37). Thereby, most studies examining the kinematics of the snatch have primarily focused on aspects of the lifters’ lower limb movements in the sagittal plane, as these are essential key variables in the execution of the snatch movement (13). Despite variations regarding the kinematics of the snatch skill level, weight category, sex (14) and a changing lifting technique over time (13), a common movement pattern has been observed in most lifters. Thereby, the snatch can be broken down into six positions and five phases (First Pull, Transition Phase, Second Pull, Turnover, Recovery) defined by changes in the knee extension angle (Figure 1) (13,31). Only a few researchers included the kinematics of the upper limbs (23, 24). Furthermore, analyzing barbell kinematics, including parameters like vertical and horizontal displacement, velocity, and acceleration, is essential for understanding efficient snatch techniques (3, 10). Weightlifting has ideal requirements for an instrumented biomechanical analysis, as all athletes perform the same movements at the same location during competitions (13). However, despite their high accuracy, MB approaches have limitations. These systems require numerous reflecting markers to be placed on bony landmarks of the human body; the preparation of the participants is very time consuming and can affect the natural motion pattern (6). Furthermore, the accuracy of MB systems relies on adequate marker placement, which is contingent upon skill and experience in physical palpation (17, 34). The accuracy of MB measurements can also be affected by soft tissue artifacts (21, 22). Thus, MB systems are difficult to use for in-field activities and competitions (27). As a result, markerless video-based (VB) motion capture techniques using synchronized video data combined with discriminative, generative, or machine learning algorithms have gained popularity in clinical and sport biomechanics settings in recent years (11, 18, 26). Human motion capture has been significantly impacted by the emergence of neural network-based methods utilizing deep learning algorithms. These systems leverage training datasets to estimate joint center positions and other salient landmarks, enabling fast and effortless motion capturing while maintaining reliability and natural motion pattern of a participant’s movement (5, 16). Thus, VB systems may introduce new possibilities in assessing kinematics during weightlifting training and competitions. One rapidly evolving VB commercial human pose estimation software is Theia3D (Theia Markerless Inc., Kingston, ON, Canada). It is a machine deep learning algorithm-based software that relies on 2D videos of several synchronized cameras and estimates the 3D pose of a human during recorded movements (35). The underlying principle of this software is described in detail in the study of Kanko et al. (17). Recent studies investigated the accuracy of the Theia3D algorithm by comparing the kinematic outcomes of lower extremity joints during gait (17) and treadmill running (38) with an MB system. The results of these studies demonstrated comparable outcomes regarding the estimated joint center positions of the two systems, which were found to be below 4 cm for all joints and showed good test-retest reliability in a laboratory setting (18). The only study assessing the kinematics of the upper limb joint angles was conducted by Lahkar et al. (21), who reported RMSDs ranging from 6.3° at the shoulder joint to 23° at the elbow joint for boxing. However, the performance of Theia3D has not been tested for overhead movements, which are challenging due to the complex anatomy of the shoulder joint with its many degrees of freedom (32). Further, the applicability and accuracy of VB systems in assessing highly dynamic full-body activities, such as the snatch, remain mostly unexplored. Hence, the aim of this study was to assess the comparability of model outputs between the VB approach (use of Theia3D) and the MB reference system (use of Vicon), regarding the kinematics of the upper and lower limbs during the snatch. Specifically, we aimed to evaluate the VB system’s applicability for technical weightlifting analyses. METHODS Participants A power analysis conducted with G*Power 3.1.9.2 (University of Kiel, Germany) determined that at least 27 trials were needed in the present study for a power of 0.80, with an effect size of 0.6 and an α=0.05 (38). Twenty-four weightlifters, competing either in the regional or national weightlifting league in Germany, volunteered to participate in this study. The inclusion criteria were at least six months of weightlifting training experience and no serious injuries within the last year. The datasets of three participants had to be discarded since neither the VB nor the MB system observed knee flexion during their transition phase, making it impossible to classify their snatch phases. Table 1 presents the characteristics of the remaining 21 participants (15 Male, 6 Female). This research was done in accordance with the ethical standards of the International Journal of Exercise Science (28). Protocol Experimental setup The VB motion capture system consisted of twelve video cameras (Baumer VLXT-31C.I.JP; 3.1-megapixel resolution) positioned approximately 2.5 m above and around the platform where the participants performed the snatch (Figure 2). The cameras recorded at 100 Hz, and video data were synchronized using Templo software (v2022.0, Contemplas GmbH, KE, Germany). To optimize video quality, settings like lighting and shutter speed (1/600 s) were constant across trials. In addition, before each test session, cameras were calibrated according to manufacturer guidelines. In addition to the VB system, an MB motion capture system (Nexus, 2.14, Vicon, OXF, UK) consisting of two Bonita 10 cameras (1-megapixel resolution), six Vantage 5 cameras (5-megapixel resolution) and two Vero v2.2 cameras (2.2-megapixel resolution) recording at 200 Hz was positioned around the lifting platform (Figure 2). These cameras were elevated to the same height as the VB cameras. We ensured that both motion capture systems shared the same global coordinate system origin and axes. At the beginning of each trial, an electrical trigger was used to synchronize both systems. Procedure All participants were familiarized with the test protocol before the respective measurement session. In addition, demographics, health history, one-repetition maximum (1RM) of the snatch, weightlifting experience, and anthropometric measures were collected by the principal investigator. Reflective markers (14 mm in diameter) were placed on the participant according to the Vicon Plug-in Gait full body model (37). Two additional markers were placed bilaterally at the iliac crest of the participants to reconstruct obstructed hip markers during the early first pull and late catch phase. Three additional reflective markers were attached to the barbell. One in the center point on each side of the barbell, and one positioned approximately three-quarters along the right side of the barbell where it did not interfere with the snatch movement. After placing the markers, participants completed an individual, self-selected warm-up. The only predetermined content of the warm-up was performing a snatch with a load of up to 50% of their self-reported 1RM. Afterward, a static calibration trial for the MB system was done. Subsequently, participants performed 2–3 repetitions at 65%, 75%, and 80% of their 1RM with two minutes of rest between repetitions. A trial was considered valid if the movement met the international competition guidelines, subjectively verified by the principal investigator (8). Multibody models and kinematic estimation: Synchronized VB data was processed using Theia3D (v2022.1.0.2309), employing the default inverse kinematics solution to obtain the 3D subject pose estimates for the limb segments. The lower body kinematic chain was defined to have three degrees of freedom (DOFs) at the hip, two DOFs at the knee (restricted in external/internal rotation), and three DOFs at the ankle. The joints of the upper extremity chain comprised three rotational DOFs at the shoulder, two DOFs at the elbow (restricted in abduction/adduction), and two DOFs at the wrist (restricted in external/internal rotation). The MB model was constrained to have three DOFs at the hip, knee, and ankle joints. For the upper extremity joints, the Plug-in Gait full body model defines three DOFs at the shoulder, one DOF at the elbow (extension/flexion), and three DOFs at the wrist. The MB vs. VB approach was evaluated by comparing estimated kinematic variables, including joint center positions of the lower and upper limbs. The MB joint center positions were calculated by the VICON Nexus software. Moreover, joint angles of the MB and VB lower limbs were compared along the sagittal and frontal plane. The joint angles of the shoulder and elbow were compared across all available DOFs between the two systems. The maximum attained extension angle and angular velocity of the weightlifters in the sagittal plane for the hip, knee, and ankle joint during the second pull, where the athletes reach full extension in these joints to displace the bar overhead (13), were calculated for the VB and MB model and examined for differences. A custom-written MATLAB (R2023a, Mathworks, Natick, MA, USA) script processed VB joint center positions and segment orientations to calculate joint angles and joint center trajectories. The utilized script applies ZYX-cardan sequence, equivalent to the procedure described by Theia3D in their documentation. Joint angles were determined using a Cardan sequence equivalent to the corresponding joint coordinate system (10). All raw kinematic trajectories of the MB system were filtered using a 4th-order zero-lag Butterworth filter with a cut-off frequency of 6 Hz. Kinematic data of both systems were exported to MATLAB, where the same filter as for the MB data was applied to the kinematic data of the VB system. Additionally, the frequency of the MB system was resampled from 200 Hz to 100 Hz. The non-time-normalized data was used to obtain the angular velocity of the joint angles. After temporal alignment through the trigger and dividing the snatch into its five phases (Figure 1), joint angles and joint center positions for the entire snatch were time-normalized for every trial. The joint angles were additionally time-normalized for each phase. The resulting kinematics of both systems were further analyzed in MATLAB. Defining the phases of the snatch To divide the snatch into its five phases (13), the kinematic data of the barbell and changes in sagittal knee angle were used to define the beginning or end of each phase. Events related to the knee kinematics were independently identified for the MB and VB model, with separate determinations for the right and left knee. The first vertical displacement of the barbell determined the initiation of the snatch. This event was solely identified by the MB system and set as the starting frame for both time-aligned datasets. The transition phase subsequently began with the knees shifting from extension to flexion. This phase was completed when the weightlifters reached the early maximum knee flexion angle. The upcoming second pull ended once the participants reached the second maximum knee extension angle. Following the second pull, the turnover phase ended when the knees of the lifters started to extend again after stabilizing the bar in the catch position (Figure 1). The first frame in which the participants achieved 98% of the maximum barbell height after rising from the overhead squat position was specified as being the end of the recovery phase, simultaneously marking the completion of the snatch for both systems. Statistical Analysis To evaluate the accuracy of Theia3D in estimating the positions of the joint centers, root mean square (RMS) values were bilaterally obtained for each participant, kinematic variable, and trial. Differences between joint angles estimated by the VB and MB system were calculated for both sides of the body across all participants, phases, and trials as root mean square differences (RMSD). The RMS values and RMSDs of the kinematic variables were averaged across the right and left sides for all recorded trials and summarized using mean values and standard deviation. In the context of a kinematic weightlifting analysis, the literature has primarily focused on the sagittal plane (13). In addition, the transverse plane outputs for the Vicon Plug-in-Gait model are notoriously unreliable and therefore, must be used with caution as the gold standard (36). Consequently, the results of the transversal plane are not presented in this study. To further identify any differences between the entire kinematic waveforms of the joint angles obtained by both systems, Statistical Parametric Mapping (SPM) from the SPM1D package (v.M.0.4.10) available for MATLAB was performed for each set of angular kinematic time-series variables, as well as for the comparison of the VB wrist and MB barbell kinematics (33). All variables were assessed for normality using a one-sample-Kolmogorov-Smirnov test with an alpha level of 0.05. A two-tailed paired t-test and Bland-Altman analysis (4) was conducted to test the differences between maximum extension angles and velocities of the lower limb joints during the second pull between the systems. Measurement errors between 2° and 5° are considered acceptable (26). For all conducted tests, statistical significance was set to α ≤ 0.05. Interpretation of Cohen’s d were interpreted as small (d = 0.2), medium (d = 0.5), and large (d = 0.8). Statistical analysis was done via custom-written MATLAB scripts and JASP (v. 0.17.2.1). RESULTS Data from 21 participants were included in the analysis. Out of the 145 snatches, 9.7% (N = 14) were discarded since the VB system could not detect knee flexion during the transition phase where the MB system captured the flexion. The remaining 131 trials were used for analysis. Joint center position estimates The VB system exhibited higher accuracy in estimating the joint center positions of the lower limbs (RMS: 4.7 ± 1.2 cm) compared to the upper limbs (RMS: 5.7 ± 1.5 cm). Among the lower limb joint centers, the ankle exhibited the smallest 3D Euclidean distance between the corresponding VB and MB joint centers (RMS: 3.5 ± 0.7 cm), followed by the knee (RMS: 4.7 ± 1.5 cm), and the hip (RMS: 6.0 ± 1.3 cm). Within the upper limb joint centers, the elbow joint center demonstrated the highest level of agreement between the two systems (RMS: 5.1 ± 1.9 cm), followed by the wrist (RMS: 5.7 ± 0.9 cm), and the shoulder (RMS: 6.3 ± 1.6 cm). The SPM analysis comparing the trajectory of the wrist, measured by the VB system, and the barbell, measured by the MB system, showed significant differences throughout the majority of the movement along the anterior-posterior direction. Furthermore, the SPM analysis revealed significant differences over the entire movement along the superior-inferior direction (Figure 3). The estimated height of the VB wrist was consistently significantly lower (i.e., inferior) than the barbell. Joint angles: RMSD values of the joint angles for the lower extremities are shown in Table 2. The data did not allow a comparison of the shoulder kinematics since the underlying kinematic models of the two systems differ regarding their rotation sequence and reference position. Among the different phases, the second pull exhibited lowest RMSD values across all DOFs of the lower limbs (RMSD: 10.1 ± 4.8°), followed by the turnover phase (RMSD: 11.0 ± 5.7°), recovery phase (RMSD: 11.6 ± 5.9°), transition phase (RMSD: 12.3 ± 5.8°), and the first pull (RMSD: 13.0 ± 6.4°). In terms of comparing across DOFs of the lower extremities, the lowest RMSDs throughout the snatch were observed along the frontal plane (RMSD: 11.2 ± 5.9°), followed by the sagittal plane, where an RMSD of 13.6 ± 4.7° was seen. Along the sagittal plane, the highest agreement between the two systems was observed at the knee (RMSD: 8.3 ± 1.9°) and the lowest at the hip (RMSD: 22.5 ± 7.0°). The VB estimates showed an offset towards knee extension (Figure 4). Additional data and Figures to the SPM analysis of the snatch divided into its five phases for all DOFs can be found in the Supplementary Material. A similar but larger offset was also observed for the hip and ankle angle (RMSD: 10.0 ± 5.2°) in that plane (Figure 5). The hip abduction/adduction demonstrated the most comparable results between the two systems (RMSD: 6.6 ± 2.9°), whereas the knee showed the highest differences in the frontal plane with an RMSD of 18.2 ± 9.0°. The VB system estimated a smaller abduction angle at the hip with a similar waveform as the MB system, but a constant offset (Figure 5). The MB ankle eversion/inversion angle (RMSD: 9.0 ± 5.7°) showed a higher variability across the trials compared to the VB counterpart (Figure 5). An overview of all RMSD values for each DOF throughout the whole snatch and each phase is presented in Table 2. Maximum extension angles and angular velocities during the second pull: Differences and the agreement of the maximum extension angles and angular velocities of the weightlifters’ lower extremities during the second pull in the sagittal plane, as estimated by both systems, are presented in Figure 6. Significant differences between the two systems were observed for all investigated parameters. The VB system displayed a significantly larger maximum extension angle at the hip, knee, and ankle joint during the second pull, with a large, medium, and small effect size, respectively. However, maximum angular extension velocities of the lower limb joints were significantly greater for the MB system and demonstrated a large effect size. DISCUSSION The aim of this study was to assess the accuracy of the VB motion capture software Theia3D in obtaining the kinematics of the lower and upper limb joints during the snatch. To achieve this, joint center positions, joint angles, maximum extension angles, and angular velocities were measured and compared to the MB motion capture system estimates. The Vicon Plug-in Gait full body model was used for the MB system. Stief et al. (36) found high inter-trial and inter-session reliability for the hip, knee, and ankle kinematic parameters in the sagittal and frontal plane but poor reliability for the knee in the frontal plane for this model. However, it is a widely used model in weightlifting analysis (37) as the position of the markers on the body ensure that the snatch can be carried out without markers being in the way of the barbell, which would make an analysis impossible. Furthermore, movements in the transverse plane are of negligible importance in the context of weightlifting analysis, as it is still mostly unexplored (13). Across all recorded trials, the differences in estimated joint center positions between the two systems were below 5 cm for the ankle and knee joint centers and below 6 cm for the hip joint center. The disparities observed in the current study for the lower limb joint centers were higher (range: 3.5 – 5.7 cm) than the ones that Kanko et al. (17,20) found during treadmill walking (range: 2.4 – 3.6 cm) and treadmill running (range: 2.1 – 3.0 cm). However, the researchers also observed the greatest differences at the hip joint. It is worth noting that Kainz et al. (16) reported errors of up to 5.08 cm for MB joint center location techniques. In contrast to Lahkar et al. (21), who found differences in a range of 1.8 cm – 3.1 cm for the upper limb joint centers during boxing and Kanko et al. (17), who reported differences between 1.1 cm and 2.4 cm, all corresponding upper limb joint centers in this study had average 3D Euclidean distances greater than 5 cm. The larger differences in joint center positions observed in this study between the VB and MB estimates of the lower and upper extremities compared to previous studies may be due to occlusion caused by the weights and the barbell. However, it should be noted that the exact anatomical landmarks and training images used by the Theia3D software are not publicly available, making this explanation speculative. Regarding weightlifting techniques, the analysis of barbell kinematics is important as it involves weightlifters’ physical movements and is the result of the forces that the lifter applies to it (3). Variable parameters, such as the vertical and horizontal displacement of the barbell, its velocity, and its acceleration, have been reported to be hallmarks of an efficient snatch technique (11). The SPM comparing the trajectories of the wrist and barbell in the sagittal and transverse plane showed that the kinematics of the VB wrist joint center cannot be used to capture the ones of the barbell. Therefore, if coaches, athletes, or researchers aim to obtain the barbell kinematics, it is advisable to position additional cameras bilaterally tracking its motion (3, 8). While there is no universally accepted error threshold for weightlifting movements, McGinley et al. (26) suggested that errors of up to 5° are generally acceptable for clinical gait analysis. In the present study, almost all estimated lower limb joint angles exceeded this value (Table 2, Range: 4.7 – 29.9°), which is also indicated by the SPM results that demonstrated significant differences between the two systems throughout a majority of the snatch movement for most DOFs. Furthermore, the VB system failed to detect knee flexion during the transition phase in 9.7% of the recorded trials where the MB system captured the flexion, highlighting its limitations. In the sagittal plane, the VB system captured comparable waveforms for all lower limb joints (Figure 5). The hip and ankle joints consistently displayed an offset towards extension, which was also evident in the knee joint during most of the snatch movement. These findings are consistent with the results of Ito et al. (15) during the squat. Additionally, Song et al. (35) observed a similar trend at the hip joint across various movements within the sagittal plane (RMSD: 6.7° – 13.8°), while Kanko et al. (17) reported a comparable trend for the sagittal ankle angle during walking (RMSD: 6.7°). The performance of pose estimation is task specific (20, 26, 28), indicating that the higher RMSD values observed in this study may be due to weaker accuracy in estimating the segment poses during the highly dynamic snatch. Moreover, the VB and MB hip joint angles displayed similar waveforms for the frontal plane, which is in line with previous findings of Kanko et al. (17) and Song et al. (35). These offsets are likely due to distinct definitions of pelvis orientation between the two systems (17,35). Ankle eversion/inversion showed the biggest differences between both systems concerning the variability. This could be attributed to the predominantly off-sagittal motions during the snatch, which were shown to be susceptible to skin artifacts (1). Therefore, the estimated angular kinematics in the frontal plane should be treated with caution. Regarding the upper limb joint angles, making a definite assessment of Theia3D’s performance within this study presents challenges. Despite the high accuracy typically associated with MB systems, issues in tracking shoulder movements have been reported, as the shoulder is a complex joint that is connected to the scapular girdle and therefore has the ability to move in several planes simultaneously (32). Given that both systems employ distinct approaches for calculating shoulder angles, coupled with the inherent difficulty of distinguishing the plane of motion, it is not evident whether either system accurately measured shoulder kinematics. Moreover, it remains uncertain which, if any, of the two systems provided an accurate depiction of shoulder movement. The MB elbow joint showed a maximum extension angle of approximately 160°, despite the weightlifters holding the barbell with nearly fully extended arms from the catch position onward, suggesting that the MB system could not accurately estimate elbow extension/flexion. Therefore, a meaningful comparison between the two systems regarding elbow extension/flexion became impossible. Despite displaying greater extension angles, the VB system captured substantially lower angular extension velocities for all lower limb joints during the second pull (Table 3). The VB system recorded a maximum angular extension velocity of 290.25 ± 34.48°·s-1 for the hip, 397.43 ± 56.34°·s-1 for the knee, and 331.64 ± 69.39°·s-1 for the ankle. In contrast, the MB system estimated angular velocities ranging from 458.0°·s-1 to 472.9°·s-1, with the hip exhibiting the highest velocity. The MB estimates of the hip and knee joint velocity agree well with the findings of Baumann et al. (3), who reported that the extension at the hip occurs faster than at the knee in the second pull, indicating that the VB system was not able to estimate the true extension velocity for the hip or the knee. The Bland and Altman analyses further indicate that the discrepancies between the MB and VB system tend to augment with increasing maximum angles and angular velocities (Figure 6). The lower maximum hip angular extension velocity estimated with the VB system may be due to the consistent offset that decreased throughout the first three phases, consequently leading to lower amplitude throughout the first three phases (Figure 7). Gourgoulis et al. (9) and Akkus (2), who investigated professional athletes, found maximum angular extension velocities at the ankle joint of 234°·s-1 to 366°·s-1 and 341°·s-1, respectively. Considering that all the mentioned studies investigated angular velocities of elite weightlifters lifting weights corresponding to their 1RM, which consequently leads to lower angular extension velocities compared to submaximal loads (11), the observed ankle extension velocity captured by the MB system in this study falls within a plausible range. However, while the angular extension velocity of the VB ankle joint goes in line with findings from the mentioned studies, it is important to note that assessing foot kinematics would be improved by capturing the movement barefoot (30), which is impractical within the context of weightlifting. Consequently, a reliable statement regarding the accuracy of the systems cannot be made. Data collection and setting up the setup was far less time-consuming and more intuitive for the VB system than the MB system. However, the post-processing time, which is dependent on the computer specifications, was higher for the VB system. Hence, unlike the MB system, providing direct feedback to athletes using the VB system is so far impossible. Consequently, the scope for integrating Theia3D into training is limited and cannot replace the traditional 2D approach derived from a single camera setup. Moreover, a notable limitation of the VB system concerning weightlifting analysis lies in its incapability to record barbell kinematics, which is a critical and widely used performance parameter (13). Furthermore, once the data is collected and processed by Theia3D, no adjustments can be made, unlike MB systems, where modifications are possible (15). Several potential error sources may have influenced the VB and MB measurements in this study. One important limitation is the difference in biomechanical model definitions between the VB and MB systems, which may have had a non-negligible impact on the comparability of motion analysis (40). This discrepancy arises from the differing methodologies employed by the VB software, which utilizes inverse kinematics for estimations, and Vicon which employs a direct kinematics approach to calculate joint model parameters. Consequently, this variance may have resulted in divergent model outputs. The most significant impact of different models was observed at the shoulder joint, where varying rotation sequences and differing reference coordinate systems made it impossible to compare the outputs for the shoulder joint. When comparing shoulder movement between systems an approach for future research might be the use of quaternions (7). Moreover, this study only assessed joint angles and positions, although information about body segments and their orientation could provide additional insights into performance parameters. Since the weightlifting technique is highly influenced by the anthropometric characteristics of an athlete (39), future studies should consider including segment pose measurements in their analysis. As for the VB system, its kinematics could have been affected by factors associated with the implemented algorithm, including potential omissions or biases within the training datasets, leading to errors in situations where the training was weak (17). The current study was the first to evaluate the 3D joint center positions and joint angles of the lower and upper extremities of weightlifters performing the snatch using both a VB and MB approach. When comparing the estimated joint center positions of the two systems, the lower extremities showed smaller differences than the upper extremities. The VB system failed to detect knee flexion during the transition phase for ~10% of the trials where the MB system captured the flexion. Furthermore, Theia3D is not able to capture barbell kinematics. Analysis of the lower limb joint angles revealed the highest level of agreement in the frontal plane, followed by the sagittal plane. For the latter, all three VB joint angles showed similar waveforms as their MB counterpart with an offset towards extension. Shoulder kinematics exhibited large disparities. The SPM analysis displayed significant differences throughout most of the snatch for all DOFs. Maximum angles and angular velocities differed substantially between the VB and MB systems. The VB motion capture system has shown its potential to overcome certain practical limitations of the MB systems and enable coaches and athletes to obtain 3D kinematics during weightlifting training and competition. However, our data suggests a limited comparability of the kinematic data output of both systems, which makes it impossible to draw reliable conclusions about the weightlifters’ technique by obtaining the kinematics with the VB system, since most previous 3D snatch kinematic data was measured using a MB system and its underlying models. Thus, it is important to acknowledge that, although the MB system is generally regarded as the reference device, it does not represent the absolute ground truth in this study due to the different biomechanical models and assumptions employed by the two systems. However, given the rapid advancements of neural network-based approaches, the VB software holds the potential to become a suitable alternative to MB approaches in weightlifting analysis in the future. Supplementary Information