Electromagnetic trackers locate small sensor coils by measuring the magnetic field from a nearby emitter. They need no line of sight, which makes them attractive for capturing human motion under clothing or for guiding needles inside tissue. Their weak point is metal. Steel beams, operating tables, and the electronics inside an ultrasound probe bend the field, and the tracker reports the wrong position. Today this is handled in one of three ways: calibrate the room in advance, throw away readings that look corrupted, or train a network to absorb the error. We propose a fourth option. Treat the distortion as an unknown, spatially local correction field. Estimate it online, together with the body pose, from the redundancy that a body model already provides: bone lengths do not change, and each sensor's offset from the skin is fixed. Then solve the resulting joint problem with a second-order method, meaning one that uses curvature information as Newton's method does, and that exploits the fact that each reading touches only a few correction coefficients.
The second half of the claim is about compute. A self-calibrating tracker is only useful if it runs on the body-worn base station or a phone. The literature reports sparse second-order speedups as wall-clock numbers measured on desktop machines. Whether exploiting sparsity actually beats a plain dense solve at the problem sizes this application produces is a hardware question, and it may come out differently on an embedded processor. We therefore treat "sparse versus dense" as a crossover point to be measured on several platforms, not as a premise.
Core Claim
Body-worn electromagnetic tracking has the same structure as bundle adjustment, the joint refinement of camera poses and 3D points in visual mapping. The distortion coefficients play the role of map points. Each has local support, so a sensor reading at a given location touches only the handful of coefficients whose basis functions cover that spot. Per-frame body poses play the role of camera poses. The curvature matrix of the joint least-squares problem is therefore arrow-shaped: there is a large block-sparse region linking coefficients to frames and no direct coupling between distant coefficients. The Schur complement, an exact way to eliminate one block of variables and leave a smaller system in the rest, then reduces each Gauss-Newton step to a system over only the coefficients the recent motion actually touched. We claim three things. First, this joint estimate reduces joint position error under real metallic distortion compared with uncorrected fitting and with the learned fitter of EM-POSE. Second, restricting the solve to the "active set" of touched coefficients keeps the per-frame cost bounded as the correction field grows. Third, the problem size at which the sparse solver overtakes a dense one depends measurably on hardware, and hardware-independent counters predict that point better than desktop timings do.
Concretely, let $y_{t,i}$ be the pose that sensor $i$ reports at frame $t$, let $g_i(\theta_t)$ be the pose the body model with joint parameters $\theta_t$ predicts for that sensor, and let $\phi_k$ be compactly supported basis functions over the sensor's position relative to the emitter, with unknown coefficients $c_k$. We minimize
$$ \sum_{t,i} \Big\| y_{t,i} - g_i(\theta_t) - \sum_k \phi_k\big(g_i(\theta_t)\big)\, c_k \Big\|^2_{\Sigma} + \lambda \|D c\|^2 + R(\theta), $$
where $\Sigma$ is the measurement covariance, $D$ is a smoothness operator over neighbouring coefficients, and $R$ is the body-model prior. Only coefficients with $\phi_k \neq 0$ near some observed sensor enter a given step. This is the source of the sparsity, and it is also the active set that shifts as the person moves or as a metal object enters the volume. We work at the level of reported 6D poses rather than raw coil voltages, because that is what public datasets provide. This choice also lets the method attach to commercial trackers without access to their internal inversion.
Relation to Prior Work
Two papers motivate the gap. EM-POSE fits the SMPL body model to 6–12 wearable electromagnetic sensors with Learned Gradient Descent, a network trained to predict update steps. It reports 31.8 mm mean per-joint position error (MPJPE) with 12 sensors, and it notes that metal within about 1.5 m degrades accuracy. It models distortion only implicitly through training data and uses no explicit second-order solve. EMDB runs a multi-stage optimization over electromagnetic, video and phone-odometry data and acknowledges localized distortion from metal structures. It does not estimate a correction field or analyze the structure of its own problem.
Two papers supply the method. SchurVINS uses the Schur complement inside a Kalman filter to eliminate visual landmarks. It reports 0.075 m trajectory error on EuRoC at 3.83 ms per frame on MH01. We borrow the elimination, but our eliminated block is a physical correction field rather than visual landmarks, and its support changes with motion. SGN rewrites dense Gauss-Newton systems as sparse saddle-point systems. It reports that the rewrite can break even at "a few hundred parameters in some cases". That is roughly our regime, and it is the reason we treat the crossover as an open, measurable quantity. SCT detects sparsity patterns by tracing which inputs affect which outputs. It distinguishes local patterns (valid only at the current point) from global ones. We use the local pattern to recompute the active set cheaply when it shifts.
One paper constrains the design. MDR stores distortion as fixed voxel "anchors" (0.1 m cubes) for magnetometer orientation. It cuts orientation error to 17.73° from MUSE's 27.42°, but it assumes a static field, and its particle filter consumes about 87% of compute. Our correction field is likewise spatially local. Unlike MDR's anchors, it is re-estimated online from kinematic consistency and solved by Gauss-Newton rather than sampling. The particle-filter cost is exactly what we aim to avoid.
On evaluation, Predicting SAT Solver Performance on Heterogeneous Hardware shows that solver runtimes shift across 21 machines and can be partially predicted from machine features. ProS measures progress with a deterministic operation count instead of seconds. We adopt both ideas for linear solvers.
Research Program
We build the joint estimator in three layers. The first layer is the static problem: a batch solve over a recorded sequence, comparing dense Cholesky on the full system against Schur elimination of the coefficient block. On the same data, both should reach the same optimum, so any difference between them is purely computational. The second layer is a sliding-window version. Coefficients outside the active set are frozen, with their information carried forward as a prior, and only touched coefficients enter the solve. When an incoming metal object shifts the distortion, large residuals concentrated in a region should reactivate and relax the prior on the coefficients there. The third layer is the hardware study. We run every solver variant on three classes of machine: a desktop x86 CPU, an ARM single-board computer comparable to a body-worn base station, and a phone-class system-on-chip. For each we record wall-clock time distributions alongside counts that do not depend on hardware: nonzeros after factorization, floating-point operations, and solver iterations. Following the regression approach of the SAT study, we then test which counters predict per-platform runtime and where the sparse-versus-dense crossover falls on each machine.
A real risk is that the distortion is not identifiable. A smooth correction field could be partly absorbed by body shape or by the sensor-to-skin offsets. We address this directly. The smoothness weight $\lambda$, fixed bone lengths, and per-subject offsets estimated once in a clean environment break the ambiguity. Before any accuracy claim, we inspect the eigenvalues of the reduced curvature matrix to identify which directions of the field are actually constrained.
Baselines
The naive reference is the same Gauss-Newton body-model fit run on uncorrected sensor poses, with no distortion terms. It shares every component except the correction field, so its gap to our method measures what the correction buys. We run it with the identical body prior and the same number of iterations.
The strongest existing method is the Learned Gradient Descent fitter of EM-POSE. It reports 31.8 mm MPJPE with 12 sensors and 35.4 mm with 6 sensors on the EM-POSE dataset. We use the authors' training protocol and evaluate on the same splits. Matching its reported clean-condition error would be acceptable. Beating it means lower MPJPE on sequences with added or recorded metal distortion, with no loss on clean sequences.
A static pre-mapped correction, in the spirit of MDR, is the natural competitor for the "learn the field" ingredient. We fit the same basis offline on a calibration recording, freeze it, and use it at test time. Our method beats it if it does better after the metal configuration changes between calibration and test.
A robust-loss variant down-weights readings with large residuals instead of correcting them. It is the optimization analogue of the "turn off the magnetometer" strategy that MDR improves upon.
The key ablation is the same joint model solved with a dense factorization. It separates accuracy gains, which should be identical, from computational gains. Our solver beats it only on platforms and at field sizes where its measured time is lower. A per-frame cost that stays flat as the field grows would additionally demonstrate the value of the active set. Every variant gets the same windows, iteration caps, and tuning budget, with $\lambda$ and window length chosen on a held-out validation split.
Why not just…?
Why not just use dense Gauss-Newton, since the problem is small? This is a real threat. Twelve sensors plus a modest field may be only a few hundred unknowns, which is exactly where SGN reports break-even. If the measured crossover on the embedded board lies above the field sizes that accuracy actually requires, we will report that the sparsity is unnecessary for this application. The accuracy contribution (online self-calibration) stands regardless. The hardware study is designed to settle this rather than assume it.
Why not just train the learned fitter with simulated distortion? Training-time augmentation could teach EM-POSE's fitter to absorb distortion. EM-POSE already needed large synthetic pre-training to bridge the gap between simulation and reality, and a learned correction has no means of adapting to a metal configuration it never saw. We test this by training the learned fitter with dipole-model distortion augmentation and evaluating on held-out real configurations that differ in metal type and placement. If it matches our method there, the explicit field is not worth its cost.
Why not just discard or down-weight suspicious readings? With 6–12 sensors, dropping a limb's sensor removes most of the information about that limb. MDR's comparison against distortion-avoiding MUSE (17.73° versus 27.42°) suggests that modelling distortion beats ignoring it, though that result comes from magnetometers, not active trackers. Our robust-loss baseline tests whether the same holds here.
Why not just report FLOPs? Sparse factorizations are typically limited by memory traffic rather than arithmetic, so FLOP counts can rank solvers wrongly. Rather than assert this, we measure it: the regression across platforms tells us whether FLOPs, factor nonzeros, or neither predicts the runtime ordering.
Proposed Experiments
We evaluate on three data sources. The first is the EM-POSE dataset (45 sequences, about 36.8 minutes, 5 subjects, with optical ground truth), both as recorded and with synthetic distortion injected from dipole models of steel and aluminium objects at controlled distances. The second is EMDB (81 sequences, 105,000 frames), where the uncorrected sensor data is used to check whether the estimated field reduces kinematic inconsistency. EMDB reports 2.3 cm positional and 10.6° angular error for its own pipeline. The third is a small new recording with a wearable rig and optical ground truth, in which metal plates of known alloy are placed near the participant and then moved mid-sequence. This recording provides a held-out real-distortion test and a dynamic intruder test.
Accuracy is measured as MPJPE and per-sensor position and angle error. The dynamic case is assessed by recovery time after the intruder moves. Compute is measured as per-frame latency distributions on each platform, plus the hardware-independent counts above.
Three checks decide the outcome. The first is whether joint estimation lowers MPJPE on the held-out real-metal recordings relative to both the naive fit and the distortion-augmented learned fitter, while leaving clean-sequence error unchanged. The second is whether the windowed active-set solver recovers within a bounded number of frames after a plate moves, where the frozen pre-mapped field does not. The third is whether, on the ARM board, the sparse solver beats dense factorization at the field resolution the accuracy experiments select. If it does not, the paper's computational claim collapses to "dense is fine here", and we would say so.
References
- EM-POSE: 3D Human Pose Estimation from Sparse Electromagnetic Trackers
- EMDB: The Electromagnetic Database of Global 3D Human Pose and Shape in the Wild
- SchurVINS: Schur Complement-Based Lightweight Visual Inertial Navigation System
- SGN: Sparse Gauss-Newton for Accelerated Sensitivity Analysis
- Sparser, Better, Faster, Stronger: Sparsity Detection for Efficient Automatic Differentiation
- Magnetic Distortion Resistant Orientation Estimation
- Predicting SAT Solver Performance on Heterogeneous Hardware
- ProS: data series progressive k-NN similarity search and classification with probabilistic quality guarantees