Accurate 3D human pose estimation from single images is possible with sophisticated deep-net architectures that have been trained on very large datasets. However, this still leaves open the problem of capturing motions for which no such database exists. Manual annotation is tedious, slow, and error-prone. In this paper, we propose to replace most of the annotations by the use of multiple views, at training time only. Specifically, we train the system to predict the same pose in all views. Such a consistency constraint is necessary but not sufficient to predict accurate poses. We therefore complement it with a supervised loss aiming to predict the correct pose in a small set of labeled images, and with a regularization term that penalizes drift from initial predictions. Furthermore, we propose a method to estimate camera pose jointly with human pose, which lets us utilize multiview footage where calibration is difficult, e.g., for pan-tilt or moving handheld cameras. We demonstrate the effectiveness of our approach on established benchmarks, as well as on a new Ski dataset with rotating cameras and expert ski motion, for which annotations are truly hard to obtain.would be to annotate video data. However, achieving high accuracy would require a great deal of annotation, which is tedious, slow, and error-prone. As illustrated by Fig. 1, we therefore propose to replace most of the annotations by the use of multiple views, at training time only. Specifically, we use them to provide weak supervision and force the system to predict the same pose in all views.While such view consistency constraints increase accuracy, they are unfortunately not sufficient. For example, the network can learn to always predict the same pose, independently of the input image. To prevent this, we use a small set of images with ground-truth poses, which serve a dual purpose. First, they provide strong supervision during training. Second, they let us regularize the multi-view predictions by encouraging them to remain close to the predictions of a network trained with the scarce supervised data only.In addition, we propose to use a normalized pose distance to evaluate all losses involving poses. It disentangles pose from scale, and we found it to be key to maintain accuracy when the annotated data is scarce.Our experiments demonstrate the effectiveness of our weakly-supervised multi-view training strategy on several 1 arXiv:1803.04775v2 [cs.CV]

The aim of this study was to assess the performance of different kinematic features measured by foot-worn inertial sensors for detecting running gait temporal events (e.g., initial contact, terminal contact) in order to estimate inner-stride phases duration (e.g., contact time, flight time, swing time, step time). Forty-one healthy adults ran multiple trials on an instrumented treadmill while wearing one inertial measurement unit on the dorsum of each foot. Different algorithms for the detection of initial contact and terminal contact were proposed, evaluated and compared with a reference-threshold on the vertical ground reaction force. The minimum of the pitch angular velocity within the first and second half of a mid-swing to mid-swing cycle were identified as the most precise features for initial and terminal contact detection with an inter-trial median ± IQR precision of 2 ± 1 ms and 4 ± 2 ms respectively. Using these initial and terminal contact features, this study showed that the ground contact time, flight time, step and swing time can be estimated with an inter-trial median ± IQR bias less than 12 ± 10 ms and the a precision less than 4 ± 3 ms. Finally, this study showed that the running speed can significantly affect the biases of the estimations, suggesting that a speed-dependent correction should be applied to improve the system’s accuracy.

This study aimed to introduce and validate a new method to estimate and correct the orientation drift measured from foot-worn inertial sensors. A modified strap-down integration (MSDI) was proposed to decrease the orientation drift, which, in turn, was further compensated by estimation of the joint center acceleration (JCA) of a twosegment model of the foot. This method was designed to fit the different foot strike patterns observed in running and was validated against an optical motion-tracking system during level treadmill running at 8, 12, and 16 km/h. The sagittal and frontal plane angles obtained from the inertial sensors and the motion tracking system were compared at different moments of the ground contact phase. The results obtained from 26 runners showed that the foot orientation at mean stance was estimated with an accuracy (inter-trial median ± IQR) of 0.4 ± 3.8 • and a precision (inter-trial precision median ± IQR) of 3.0 ± 1.8 •. The orientation of the foot shortly before initial contact (IC) was estimated with an accuracy of 2.0 ± 5.9 • and a precision of 1.6 ± 1.1 • ; which is more accurate than commonly used zero-velocity update methods derived from gait analysis and not explicitly designed for running. Finally, the study presented the effect initial and terminal contact (TC) detection errors have on the orientation parameters reported.

The individualized models were capable of discriminating different techniques performed by advanced skiers and seemed more accurate than the generalized models. The models presented here offer a simple yet accurate method to estimate the aerodynamic drag acting upon alpine skiers while rapidly moving through the range of positions typical to turning technique.

This study used wireless technology to investigate joint kinematic characteristics of the four alpine skiing disciplines. Knee and hip angles were measured in 20 national team alpine skiers during 253 ski runs under FIS regulation, including: 85 Slalom (SL), 123 Giant Slalom (GS), 29 Super Giant Slalom (SG), and 16 Downhill (DH). Data were analyzed by outside (OL, n = 2,087) and inside leg (IL, n = 2,015). The proportion of concentric and eccentric phases (extension and flexion respectively for the knee extensors) as well as the proportion of the quasi-isometric phase defined between ±20 • .s −1 depended on the discipline in interaction with the IL/OL (p < 0.001). The results showed a lower knee quasi-isometric duration on OL in SL (11%) than other disciplines (DH: 38%; SG: 42%; GS: 34%, p < 0.001, d > 1.8), suggesting a highly dynamic style. Quasi-isometric mode was significantly longer on OL than IL in GS (34 vs. 20%, p < 0.001, d = 1.16) and SG (42 vs. 28%, p < 0.001, d = 1.11) but was significantly longer on IL than OL in SL (19 vs. 11%, p < 0.001, d = 0.64). Thus, GS and SG showed similarities, with a significantly faster knee eccentric mean angular velocity on IL compared to OL (GS −58 vs. −54 • .s −1 , SG −52 vs. −45 • .s −1 , p < 0.001, d ≥ 0.22) whereas SL showed an opposite pattern (−72 vs. −89 • .s −1 , p < 0.001, d = 1.10). The quasi-isometric phase was overlooked in previous studies but is crucial to consider. The current data may be used to train the outside and inside leg specificities incorporating discipline-specific contraction modes and exercises.