Multi-animal pose estimation, identification and tracking with DeepLabCut.

Estimating the pose of multiple animals is a challenging computer vision problem: frequent interactions cause occlusions and complicate the association of detected keypoints to the correct individuals, as well as having highly similar looking animals that interact more closely than in typical multi-human scenarios. To take up this challenge, we build on DeepLabCut, an open-source pose estimation toolbox, and provide high-performance animal assembly and tracking-features required for multi-animal scenarios. Furthermore, we integrate the ability to predict an animal's identity to assist tracking (in case of occlusions). We illustrate the power of this framework with four datasets varying in complexity, which we release to serve as a benchmark for future algorithm development.

Backstroke-to-Breaststroke Turns Muscular Activity. A Study Conducted in Age Group Swimmers.

The aims of this study were to compare surface electromyographic (EMG) activity and kinematic variables among open, somersault, bucket and crossover backstroke-to-breaststroke turning techniques, and identify relationships between the integrated electromyography (iEMG) and kinematics profile focusing on the rotation and push-off efficacy. Following a four-week of systematically increasing contextual interference intervention program, eight 12.38 ± 0.55 years old male swimmers randomly performed twelve repetitions (three in each technique) turns in and out of the wall at maximum speed until the 7.5 m reference mark. Surface EMG values of the right vastus lateralis, biceps femoris, tibialis anterior, gastrocnemius medialis, rectus abdominis, external oblique, erector spinae and latissimus dorsi were recorded and processed using the integrated electromyography (iEMG) and the total integrated electromyography (TiEMG) that was expressed as a percentage of iEMGmax to normalize per unit of time for each rotation and push-off phase. Complementarily, 2D sagittal views from an underwater video camera were digitized to determine rotation and push-off efficacy. The crossover turn presented the highest rotation and push-off iEMG values. Erector spinae and gastrocnemius medialis had the highest activity in the rotation and push-off phases (89 ± 10 and 98 ± 69%, respectively). TiEMG depicted a very high activity of lower limb muscles during push-off activity (222 ± 17 to 247 ± 16%). However, there were no relation between TiEMG and rotation and push-off time, tuck index and final push-off velocity during the rotation and the push-off phases across all the studied turning techniques. The rotation efficacy in age-group swimmers were dependent on rotation time (p = 0.04). The different turning techniques were not distinguishable regarding iEMG activity as a possible determinant of rotation and push-off efficacy. Our study has direct implications for selecting appropriate exercises and designing training programs for optimizing the rotation and push-off phases of backstroke-to-breaststroke turning at young ages.

Modulation of upper limb joint work and power during sculling while ballasted with varying loads.

The human musculoskeletal system must modulate work and power output in response to substantial alterations in mechanical demands associated with different tasks. In particular, in water, upper limb muscles must perform net positive work to replace the energy lost against the dissipative fluid load. Where in the upper limb are work and power developed? Is mechanical output modulated similarly at all joints, or are certain muscle groups favored? This study examined, for the first time, how work and power per stroke are distributed at the upper limb joints in seven male participants sculling while ballasted with 4, 6, 8, 10 and 12 kg. Upper limb kinematics was captured and used to animate body virtual geometry. Net wrist, elbow and shoulder joint work and power were subsequently computed through a novel approach integrating unsteady numerical fluid flow simulations and inverse dynamics modeling. Across a threefold increase in load, total work and power significantly increased from 0.38±0.09 to 0.67±0.13 J kg-1, and 0.47±0.06 to 1.14±0.16 W kg-1, respectively. Shoulder and elbow equally supplied >97% of the upper limb total work and power, coherent with the proximo-distal gradient of work performance in the limbs of terrestrial animals. Individual joint relative contributions remained constant, as observed on land during tasks necessitating no net work. The apportionment of higher work and power simultaneously at all joints in water suggests a general motor strategy of power modulation consistent across physical environments, limbs and tasks, regardless of whether or not they demand positive net work.

SuperAnimal pretrained pose estimation models for behavioral analysis.

Quantification of behavior is critical in diverse applications from neuroscience, veterinary medicine to animal conservation. A common key step for behavioral analysis is first extracting relevant keypoints on animals, known as pose estimation. However, reliable inference of poses currently requires domain knowledge and manual labeling effort to build supervised models. We present SuperAnimal, a method to develop unified foundation models that can be used on over 45 species, without additional manual labels. These models show excellent performance across six pose estimation benchmarks. We demonstrate how to fine-tune the models (if needed) on differently labeled data and provide tooling for unsupervised video adaptation to boost performance and decrease jitter across frames. If fine-tuned, SuperAnimal models are 10-100× more data efficient than prior transfer-learning-based approaches. We illustrate the utility of our models in behavioral classification and kinematic analysis. Collectively, we present a data-efficient solution for animal pose estimation.

Video-driven simulation of lower limb mechanical loading during aquatic exercises.

Understanding the mechanical demands of an exercise on the musculoskeletal system is crucial to prescribe effective training or therapeutic interventions. Yet, that knowledge is currently limited in water, mostly because of the difficulty in evaluating external resistance. Here I reconcile recent advances in 3D markerless pose and mesh estimation, biomechanical simulations, and hydrodynamic modeling, to predict lower limb mechanical loading during aquatic exercises. Simulations are driven exclusively from a single video. Fluid forces were estimated within 12.5±4.1% of the peak forces determined through computational fluid dynamics analyses, at a speed three orders of magnitude greater. In silico hip and knee resultant joint forces agreed reasonably well with in vivo instrumented implant recordings (R2=0.74) downloaded from the OrthoLoad database, both in magnitude (RMSE =251±125 N) and direction (cosine similarity = 0.92±0.09). Hip flexors, glutes, adductors, and hamstrings were the main contributors to hip joint compressive forces (40.4±12.7%, 25.6±9.7%, 14.2±4.8%, 13.0±8.2%, respectively), while knee compressive forces were mostly produced by the gastrocnemius (39.1±15.9%) and vasti (29.4±13.7%). Unlike dry-land locomotion, non-hip- and non-knee-spanning muscles provided little to no offloading effect via dynamic coupling. This noninvasive method has the potential to standardize the reporting of exercise intensity, inform the design of rehabilitation protocols and improve their reproducibility.

A Primer on Motion Capture with Deep Learning: Principles, Pitfalls, and Perspectives.

Extracting behavioral measurements non-invasively from video is stymied by the fact that it is a hard computational problem. Recent advances in deep learning have tremendously advanced our ability to predict posture directly from videos, which has quickly impacted neuroscience and biology more broadly. In this primer, we review the budding field of motion capture with deep learning. In particular, we will discuss the principles of those novel algorithms, highlight their potential as well as pitfalls for experimentalists, and provide a glimpse into the future.

Shoulder mechanical demands of slow underwater exercises in the scapular plane.

BACKGROUND: The mechanical demands of underwater shoulder exercises have only been assessed indirectly via electromyographical measurements. Yet, this is insufficient to understand all the clinical implications. The purpose of this study was to evaluate musculoskeletal system loading during slow (30°/s) scapular plane arm elevation and lowering performed in two media (air vs water) and body positions (sitting vs supine). METHODS: Eighteen participants' upper bodies were scanned and virtually animated within unsteady numerical fluid flow simulations to compute hydrodynamic forces. Together with weight, buoyancy and segment inertial parameters, these were fed into an inverse dynamics model to obtain net shoulder moments, power and work. FINDINGS: Positive mechanical work done at the shoulder was 32.4% (95% CI [29.2, 35.6]) and 25.0% [22.8, 27.2] that when performing the same movement on land, supine and sitting respectively. Arm elevation was ~2.5× less demanding sitting than supine (mean 0.012 (SD 0.018) vs mean 0.027 (SD 0.012) J·kg-1, P = 0.034). Instantaneous power was consistently positive when sitting albeit very low during elevation (0.003 W·kg-1) whereas, when supine, it was alternately negative for short period (~1.2 s) and positive (~4.8 s), peaking at levels 3× higher (0.01 W·kg-1). INTERPRETATION: Performing sitting elicited concentric muscle contractions at very low effort, which is advantageous during early rehabilitation to restore joint mobility. Exercising supine, by contrast, required rapid pre-stretch followed by concentric force production at an overall higher mechanical cost, and is therefore better suited to more advanced rehabilitation stages.

Shoulder joint kinetics and dynamics during underwater forward arm elevation.

Aquatic exercises are widely implemented into rehabilitation programs. However, both evaluating their mechanical demands on the musculoskeletal system and designing protocols to provide progressive loading are difficult tasks. This study reports for the first time shoulder joint kinetics and dynamics during underwater forward arm elevation performed at speeds ranging from 22.5 to 90°/s. Net joint moments projected onto anatomical axes of rotation, joint power, and joint work were calculated in 18 participants through a novel approach coupling numerical fluid flow simulations and inverse dynamics. Joint dynamics was revealed from the 3D angle between the joint moment and angular velocity vectors, identifying three main functions-propulsion, stabilization, and resistance. Speeds <30°/s necessitated little to no power at all, whereas peaks about 0.20 W⋅kg-1 were seen at 90°/s. As speed increased, peak moments were up to 61 × higher at 90 than at 22.5°/s, (1.82 ±â€¯0.12%BW⋅AL vs 0.03 ±â€¯0.01%BW⋅AL, P < 0.038). This was done at the expense of a substantial decrease in the joint moment contribution to joint stability though, which goes against the intuition that greater stabilization is required to protect the shoulder from increasing loads. Slow arm elevations (<30°/s) are advantageous for joint mobility gain at low mechanical solicitation, whereas the intensity at 90°/s is high enough to stimulate muscular endurance improvements. Simple predictive equations of shoulder mechanical loading are provided. They allow for easy design of progressive protocols, either for the postoperative shoulder or the conditioning of athlete targeting very specific intensity regions.

Explosive lower limb extension mechanics: An on-land vs. in-water exploratory comparison.

During a horizontal underwater push-off, performance is strongly limited by the presence of water, inducing resistances due to its dense and viscous nature. At the same time, aquatic environments offer a support to the swimmer with the hydrostatic buoyancy counteracting the effects of gravity. Squat jump is a vertical terrestrial push-off with a maximal lower limb extension limited by the gravity force, which attracts the body to the ground. Following this observation, we characterized the effects of environment (water vs. air) on the mechanical characteristics of the leg push-off. Underwater horizontal wall push-off and vertical on-land squat jumps of two local swimmers were evaluated with force plates, synchronized with a lateral camera. To better understand the resistances of the aquatic movement, a quasi-steady Computational Fluid Dynamics (CFD) analysis was performed. The force-, velocity- and power-time curves presented similarities in both environments corresponding to a proximo-distal joints organization. In water, swimmers developed a three-step explosive rise of force, which the first one mainly related to the initiation of body movement. Drag increase, which was observed from the beginning to the end of the push-off, related to the continuous increase of body velocity with high values of drag coefficient (CD) and frontal areas before take-off. Specifically, with velocity, frontal area was the main drag component to explain inter-individual differences, suggesting that the streamlined position of the lower limbs is decisive to perform an efficient push-off. This study motivates future CFD simulations under more ecological, unsteady conditions.

Upper limb joint forces and moments during underwater cyclical movements.

Sound inverse dynamics modeling is lacking in aquatic locomotion research because of the difficulty in measuring hydrodynamic forces in dynamic conditions. Here we report the successful implementation and validation of an innovative methodology crossing new computational fluid dynamics and inverse dynamics techniques to quantify upper limb joint forces and moments while moving in water. Upper limb kinematics of seven male swimmers sculling while ballasted with 4kg was recorded through underwater motion capture. Together with body scans, segment inertial properties, and hydrodynamic resistances computed from a unique dynamic mesh algorithm capable to handle large body deformations, these data were fed into an inverse dynamics model to solve for joint kinetics. Simulation validity was assessed by comparing the impulse produced by the arms, calculated by integrating vertical forces over a stroke period, to the net theoretical impulse of buoyancy and ballast forces. A resulting gap of 1.2±3.5% provided confidence in the results. Upper limb joint load was within 5% of swimmer׳s body weight, which tends to supports the use of low-load aquatic exercises to reduce joint stress. We expect this significant methodological improvement to pave the way towards deeper insights into the mechanics of aquatic movement and the establishment of practice guidelines in rehabilitation, fitness or swimming performance.

Breaststroke swimmers moderate internal work increases toward the highest stroke frequencies.

A model to predict the mechanical internal work of breaststroke swimming was designed. It allowed us to explore the frequency-internal work relationship in aquatic locomotion. Its accuracy was checked against internal work values calculated from kinematic sequences of eight participants swimming at three different self-chosen paces. Model predictions closely matched experimental data (0.58 ± 0.07 vs 0.59 ± 0.05 J kg(-1)m(-1); t(23)=-0.30, P=0.77), which was reflected in a slope of the major axis regression between measured and predicted total internal work whose 95% confidence intervals included the value of 1 (ß=0.84, [0.61, 1.07], N=24). The model shed light on swimmers ability to moderate the increase in internal work at high stroke frequencies. This strategy of energy minimization has never been observed before in humans, but is present in quadrupedal and octopedal animal locomotion. This was achieved through a reduced angular excursion of the heaviest segments (7.2 ± 2.9° and 3.6 ± 1.5° for the thighs and trunk, respectively, P<0.05) in favor of the lightest ones (8.8 ± 2.3° and 7.4 ± 1.0° for the shanks and forearms, respectively, P<0.05). A deeper understanding of the energy flow between the body segments and the environment is required to ascertain the possible dependency between internal and external work. This will prove essential to better understand swimming mechanical cost determinants and power generation in aquatic movements.

Different Muscle-Recruitment Strategies Among Elite Breaststrokers.

PURPOSE: To investigate electromyographical (EMG) profiles characterizing the lower-limb flexion-extension in an aquatic environment in high-level breaststrokers. METHODS: The 2-dimensional breaststroke kick of 1 international- and 2 national-level female swimmers was analyzed during 2 maximal 25-m swims. The activities of biceps femoris, rectus femoris, gastrocnemius, and tibialis anterior were recorded. RESULTS: The breaststroke kick was divided in 3 phases, according to the movements performed in the sagittal plane: push phase (PP) covering 27% of the total kick duration, glide phase (GP) 41%, and recovery phase (RP) 32%. Intrasubject reproducibility of the EMG and kinematics was observed from 1 stroke cycle to another. In addition, important intersubject kinematic reproducibility was noted, whereas muscle activities discriminated the subjects: The explosive PP was characterized by important muscle-activation peaks. During the recovery, muscles were likewise solicited for swimmers 1 (S1) and 2 (S2), while the lowest activities were observed during GP for S2 and swimmer 3 (S3), but not for S1, who maintained major muscle solicitations. CONCLUSIONS: The main muscle activities were observed during PP to perform powerful lower-limb extension. The most-skilled swimmer (S1) was the only 1 to solicit her muscles during GP to actively reach better streamlining. Important activation peaks during RP correspond to the limbs acting against water drag. Such differences in EMG strategies among an elite group highlight the importance of considering the muscle parameters used to effectively control the intensity of activation among the phases for a more efficient breaststroke kick.

Phase-dependence of elbow muscle coactivation in front crawl swimming.

Propulsion in swimming is achieved by complex sculling movements with elbow quasi-fixed on the antero-posterior axis to transmit forces from the hand and the forearm to the body. The purpose of this study was to investigate how elbow muscle coactivation was influenced by the front crawl stroke phases. Ten international level male swimmers performed a 200-m front crawl race-pace bout. Sagittal views were digitized frame by frame to determine the stroke phases (aquatic elbow flexion and extension, aerial elbow flexion and extension). Surface electromyograms (EMG) of the right biceps brachii and triceps brachii were recorded and processed using the integrated EMG to calculate a coactivation index (CI) for each phase. A significant effect of the phases on the CI was revealed with highest levels of coactivation during the aquatic elbow flexion and the aerial elbow extension. Swimmers stabilize the elbow joint to overcome drag during the aquatic phase, and act as a brake at the end of the recovery to replace the arm for the next stroke. The CI can provide insight into the magnitude of mechanical constraints supported by a given joint, in particular during a complex movement.