The research, led by Valentin Fernandez and colleagues at INSERM U1084 and the University of Poitiers, together with collaborators in Coimbra and Samarkand, addresses a subtle but consequential blind spot. Current evaluation of pose estimation pipelines relies almost exclusively on frame-wise spatial accuracy metrics such as root mean square error (RMSE), percentage of correct keypoints (PCK) and mean average precision (mAP). These metrics measure how closely the AI’s predicted landmark positions match human annotations in individual frames. What they fail to capture, the authors contend, is whether the resulting trajectories are reliable enough to yield trustworthy kinematic biomarkers downstream.

The problem becomes acute when those biomarkers involve derivatives. Quantities such as velocity, acceleration and jerk, the rate of change of acceleration, are computed by temporal differentiation of the reconstructed trajectories. Differentiation acts as a noise amplifier: high-frequency tracking jitter and frame-to-frame fluctuations that barely register in a spatial accuracy score can be magnified into substantial distortions of the derived movement features. Two pipelines with nearly identical localisation accuracy may therefore produce trajectories of very different quality once temporal consistency, occlusion handling and kinematic fidelity are taken into account.

To close this gap, the team developed a reliability-centred evaluation framework built around a new composite metric, the Motor Biomarker Reliability Composite Score, or MBRCS. The score integrates four complementary dimensions of trajectory reliability: spatial accuracy, temporal stability, occlusion robustness and kinematic fidelity. Each component is normalised to a scale from zero to one using robust percentile-based scaling, in which values are clipped to the fifth and ninety-fifth percentiles of the evaluated distribution to reduce sensitivity to outliers. Higher values indicate greater trajectory reliability. The authors emphasise that a normalised score near zero reflects relative positioning within the benchmark, not an absence of capability.

The weighting scheme assigns greater importance to temporal stability and kinematic fidelity, a choice made a priori because derivative-based kinematic descriptors are particularly vulnerable to temporal fluctuations and noise amplification. Sensitivity analyses using alternative weighting configurations, including uniform weighting and accuracy-prioritised schemes, produced highly consistent architecture rankings with Spearman correlation coefficients above 0.9, indicating that the composite score is not strongly dependent on the precise weights chosen. The MBRCS is intended as a comparative benchmarking tool for selecting pipelines, not as an absolute measure of performance, and computational efficiency was deliberately excluded from the score.

The framework was tested on a demanding experimental paradigm: the Montoya staircase task, in which mice retrieve food pellets from descending staircases on either side of a central platform. Adult C57BL/6 mice, twelve control animals and twelve with unilateral lesions of the primary motor cortex, were recorded from a lateral view at 120 frames per second with a resolution of 1,280 by 1,024 pixels. Around 4,800 frames from 24 behavioural recordings were manually annotated, covering landmarks including the shoulder, elbow, wrist, forearm, base of the fingers and fingertips. The task is a particularly rigorous benchmark because skilled reaching involves small-amplitude distal limb displacements, rapid temporal transitions and frequent partial occlusions caused by self-occlusion and contact with the apparatus.

Nine DeepLabCut-based convolutional neural network architectures were trained and tested under identical conditions using DeepLabCut version 3.0 with a PyTorch backend, spanning ResNet-50, ResNet-101 and ResNet-152, MobileNetV2, EfficientNet-B0 and EfficientNet-B3, and several DLCRNet-based configurations. All models were trained for 600,000 iterations on the same annotated dataset with the same split, optimisation pipeline and post-processing, so that performance differences could be attributed to architecture rather than training strategy. Predicted landmarks with a probability below 0.9 were treated as low-confidence detections; short gaps of up to five consecutive frames were reconstructed by linear interpolation, while longer interruptions, which accounted for less than 3 percent of low-confidence segments, were excluded from kinematic analysis. Trajectories were then smoothed with a Gaussian filter of sigma equal to 2 frames.

The results revealed a striking dissociation. Spatial localisation metrics remained comparatively clustered across architectures, with RMSE ranging only from 2.80 to 3.58 pixels, PCK at a 5-pixel threshold from 90.70 to 95.90 percent, and mAP from 85.63 to 91.07 percent. Yet the reliability dimensions told a very different story. Temporal stability varied substantially, and normalised jerk values, the downstream biomarker used to assess kinematic fidelity, ranged from 1.60 for the best-performing DLCRNet_ms5 to 3.85 for MobileNetV2. When all four dimensions were combined, MBRCS values spanned almost the entire normalised range, from 0.00 to 0.99, with DLCRNet_ms5 at the top and MobileNetV2 at the bottom.

Crucially, these reliability differences had functional consequences. In an external validation step deliberately excluded from the MBRCS itself, the researchers assessed whether trajectories with higher intrinsic reliability better preserved biologically meaningful differences between the control and lesion groups. Using Cohen’s d effect sizes as the primary measure of known-group validity, complemented by receiver operating characteristic analysis with logistic regression classifiers under stratified five-fold cross-validation, they found that architectures with higher temporal stability and kinematic fidelity showed greater separation of jerk distributions between groups. DLCRNet_ms5 and DLCRNet_stride32 achieved both larger effect sizes and higher AUC values, whereas MobileNetV2 showed the weakest discriminative performance, with an AUC of approximately 0.60.

The authors are careful to frame their contribution as complementary rather than substitutive. Spatial localisation accuracy remains meaningful, and the best-performing architectures in the composite score were also the most accurate localisers, but the additional dimensions exposed differences that frame-wise metrics alone could not reveal. The framework’s limitations are acknowledged: the evaluation was restricted to DeepLabCut-based architectures and a single downstream biomarker, robustness was assessed primarily under occlusion-related perturbations rather than motion blur or reduced frame rate, and variability from independent training repetitions was not evaluated. Nevertheless, the researchers argue that the methodology offers a structured, reproducible path toward benchmarking pose estimation pipelines by the quality of the kinematic information they preserve, a shift with particular relevance for studies of subtle motor impairment, longitudinal monitoring and acquisition conditions plagued by partial occlusions, where the reliability of the trajectory, not just the accuracy of the pose, determines whether a biomarker can be trusted.

Subject of Research: A reliability-centred evaluation framework for markerless pose estimation in mouse movement analysis

Article Title: From pose estimation accuracy to motor biomarker reliability: a reliability-centred evaluation framework for markerless movement analysis

Article References: Fernandez, V., Abdiyeva, K., Brot, S., Perreira de Almeida, L., Gaillard, A., & Fernandez-Maloigne, C. (2026). From pose estimation accuracy to motor biomarker reliability: a reliability-centred evaluation framework for markerless movement analysis. Frontiers in Neuroinformatics, 20, Article 1790241. https://doi.org/10.3389/fninf.2026.1790241

Image Credits: AI Generated

DOI: 10.3389/fninf.2026.1790241

Keywords: markerless pose estimation, DeepLabCut, motor biomarkers, kinematic analysis, trajectory reliability, Montoya staircase task, Normalized Jerk Score, occlusion robustness, temporal stability, neuroinformatics, movement smoothness, convolutional neural networks

Developed by Jan Hejda and colleagues and described in the journal SoftwareX, BatPose is a Python package that carries a movement recording from raw stereo video all the way to biomechanical statistics. It estimates three-dimensional human pose in real physical units, computes nine joint and trunk angles, and reports range of motion, variability, and left-right symmetry, all through a graphical interface or command-line tools that run entirely on a local CPU. The software is released under the GNU General Public License, with version 1.1.0 available on GitHub, and the team has published a preliminary validation against a professional marker-based system.

The technical pipeline follows a classical stereo-geometry route rather than an end-to-end learned model. First, the user waves a ChArUco calibration board, a checkerboard embedded with fiducial markers, in front of two cameras. The software detects the board across sampled frames, deliberately keeping views that cover different regions of the image, and then solves each camera’s internal parameters and the relative position between the two cameras using Zhang’s planar calibration method. Three lens distortion models are supported, from a standard pinhole model to wide-angle and fisheye formulations, and the program warns users when the chosen model does not match the physical lens, since a mismatch can silently warp every reconstructed coordinate.

Calibration quality is graded automatically using the root mean square reprojection error: below 0.5 pixels is rated excellent, up to 1.5 pixels is acceptable, and anything higher triggers a recommendation to recalibrate, often because the lens is wider than the model allows. The developers are careful to note that this score is an in-sample statistic computed on the calibration images themselves, not a validated bound on three-dimensional accuracy at the distance where movement is later recorded. That kind of methodological honesty runs throughout the project, which distinguishes it from many flashy pose-estimation demos.

Once the cameras are calibrated, a two-dimensional pose detector locates seventeen body keypoints in each view. The default backend is MediaPipe BlazePose, with RTMPose available as an alternative, and both outputs are mapped onto the standard COCO-17 skeleton convention. The software then triangulates each keypoint in three dimensions using the direct linear transform, the workhorse algorithm of multi-view geometry. A confidence gate rejects joints whose detections fall below a threshold in either view or whose reprojection error is too large, and a One Euro filter smooths each coordinate over time with settings that users can tune to match their frame rate and the speed of the movement being studied.

A particularly thoughtful feature is the floor world frame. By triangulating the corners of a board lying on the ground and aligning them with the Kabsch algorithm, BatPose establishes an upright coordinate frame with its origin at the board center, expressed in meters. Every recording from every session then shares the same spatial reference, which makes cross-session comparisons meaningful. From the reconstructed skeletons, the package computes bilateral knee, hip, elbow, and shoulder angles as true three-dimensional geometric segment angles, plus trunk inclination relative to vertical, and derives statistics including peak angular velocity and a bilateral symmetry index originally developed for force-platform gait analysis.

The validation against a twelve-camera Qualisys marker-based system at 300 Hz gives a first quantitative picture of what to expect. Three participants performed arm raises, biceps curls, marching, forward bends, and squats under three different stereo camera configurations. Across eight flash-synchronized recordings, the frame-weighted mean per-joint position error was 52.1 millimeters, with a median of 42.1 millimeters, and 77.3 percent of joint observations met the criteria for direct measurement. Mean absolute error for the nine angle outputs averaged 7.50 degrees, with concordance correlation coefficients between 0.910 and 0.968. Notably, performance depended strongly on camera geometry: one configuration with a wider baseline and steeper convergence angle achieved errors around 45 millimeters and over 92 percent coverage, while another dropped to under 50 percent coverage.

The team also compared angle outputs against the Vicon-derived TotalCapture dataset, finding a mean recording-level error of 9.90 degrees across five participants, though with wide limits of agreement and coverage as low as 44.6 percent in some recordings. Repeatability tests showed that repeated calibrations of a fixed rig agreed to within a few millimeters of baseline and a fraction of a degree of rotation, and that repeated detections on identical input reproduced coordinates almost exactly, with occasional large discrepancies near tracking gaps. The authors repeatedly emphasize that these are preliminary technical validations under specific tested conditions, not evidence of clinical validity or generalizability to other populations, tasks, or camera arrangements.

Those caveats matter, and the limitations section is refreshingly candid. The software matches people between the two camera views by detection order rather than by identity, which means it is designed for a single subject and may fail when multiple people cross paths. Accuracy is ultimately bounded by the quality of the two-dimensional detector and the calibration, and the system reconstructs kinematics only, not the musculoskeletal forces that pipelines built on OpenSim provide. Clothing, occlusion, and camera distance were not systematically varied in controlled experiments, and derived statistics such as peak velocity and symmetry indices cannot be assumed reliable from joint-angle agreement alone. The developers plan to add cross-view identity matching and additional pose backends in future releases.

Even with those boundaries, the implications are striking. A rehabilitation clinic, a sports team, or an ergonomics laboratory in a low-resource setting could now quantify movement asymmetries, track recovery across weeks, or analyze exercise technique with equipment costing a tiny fraction of a professional capture suite, while keeping all recordings on institutional computers rather than uploading them to a cloud service. Cached intermediate files mean users can swap detectors without recalibrating and retune reconstruction without re-running detection, and a built-in camera simulator lets students learn the workflow without any hardware at all. In a field where the gap between laboratory-grade measurement and everyday practice has persisted for decades, a compact, transparent, fully local stereo pipeline with published validation numbers is a genuinely meaningful step toward democratizing the science of human movement.

Subject of Research: Markerless stereo 3D human pose estimation and biomechanical analysis software

Article Title: BatPose: Single-subject stereo 3D human pose estimation and biomechanical analysis from synchronized cameras

Article References: Hejda, J., Sokol, M., Fischer, A. G., Radus, L. F., Leová, L., Volf, P., & Kutílek, P. (2026). BatPose: Single-subject stereo 3D human pose estimation and biomechanical analysis from synchronized cameras. SoftwareX, 36, Article 103093. https://doi.org/10.1016/j.softx.2026.103093

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103093

Keywords: BatPose, motion capture, markerless pose estimation, biomechanics, stereo vision, camera calibration, MediaPipe BlazePose, triangulation, joint angles, rehabilitation, open-source software, Qualisys validation