MoCap2GT: High-Precision SLAM Ground Truth

• MoCap2GT is a joint optimization approach that fuses MoCap and IMU data to refine ground truth trajectories for SLAM evaluation.
• It estimates an IMU-centric trajectory by jointly calibrating marker-to-IMU extrinsics, time alignment, and gravity constraints.
• The method mitigates MoCap jitter and calibration errors, achieving benchmark metrics of ATE/ARE <2 mm/0.2° and improved SLAM evaluation.

to=arxiv_search.search 北京赛车怎么json data={"query":"MoCap2GT ground truth estimator SLAM benchmarking motion capture IMU fusion", "max_results": 5, "sort_by": "submittedDate", "sort_order": "descending"} to=arxiv_search.search 微信公众号天天中彩票json data={"query":"SmartMocap joint estimation human and camera motion uncalibrated RGB cameras", "max_results": 3, "sort_by": "submittedDate", "sort_order": "descending"} MoCap2GT is a joint optimization approach for generating high-precision ground-truth trajectories for SLAM benchmarking from marker-based optical motion capture and inertial measurements from the device under test (DUT) (Shu et al., 17 Jul 2025). Its central premise is that raw MoCap output, although drift-free and routinely treated as “ground truth,” is not sufficiently accurate for modern SLAM evaluation because spatiotemporal calibration errors between the MoCap system and the DUT, together with high-frequency MoCap jitter, corrupt absolute rotation error and short-horizon relative pose metrics. MoCap2GT therefore estimates an IMU-centric trajectory in the DUT clock and body frame, while jointly refining marker-to-IMU extrinsics, time alignment, and gravity alignment.

1. Problem setting and state representation

MoCap2GT addresses the problem of recovering a trajectory that is simultaneously consistent with the DUT IMU frame and clock, globally constrained by MoCap, and suitable for evaluating ATE, ARE, RTE, and RRE in contemporary SLAM systems (Shu et al., 17 Jul 2025). The desired output is not a smoothed marker trajectory in the marker frame, but a trajectory pose $\boldsymbol{T}_{\mathrm{WI}} \in SE(3)$ centered on the IMU body frame $\mathcal{F}_{\mathrm{I}}$ and expressed in the MoCap world frame $\mathcal{F}_{\mathrm{W}}$.

The formulation explicitly distinguishes four frames: the MoCap world frame $\mathcal{F}_{\mathrm{W}}$, the marker body frame $\mathcal{F}_{\mathrm{M}}$, the gravity-aligned IMU world frame $\mathcal{F}_{\mathrm{G}}$, and the IMU body frame $\mathcal{F}_{\mathrm{I}}$ (Shu et al., 17 Jul 2025). The homogeneous transform $\boldsymbol{T}_{\mathrm{AB}}$ maps from frame $\mathcal{F}_{\mathrm{B}}$ to frame $\mathcal{F}_{\mathrm{A}}$, with rotation $\mathcal{F}_{\mathrm{I}}$0, translation $\mathcal{F}_{\mathrm{I}}$1, and quaternion $\mathcal{F}_{\mathrm{I}}$2.

The unknown state is defined as

$\mathcal{F}_{\mathrm{I}}$3

with inertial state

$\mathcal{F}_{\mathrm{I}}$4

spatiotemporal extrinsic state

$\mathcal{F}_{\mathrm{I}}$5

and world alignment state

$\mathcal{F}_{\mathrm{I}}$6

This structure encodes three facts that are fundamental to the method. First, the marker-to-IMU transform $\mathcal{F}_{\mathrm{I}}$7 is unknown. Second, the MoCap–IMU time offset is unknown and time-varying. Third, only roll and pitch are observable from gravity alignment; yaw is fixed to zero (Shu et al., 17 Jul 2025).

2. Error sources and joint maximum-likelihood estimation

MoCap2GT is motivated by two specific failure modes in conventional MoCap-based GT generation: spatiotemporal calibration errors and MoCap jitter (Shu et al., 17 Jul 2025). The paper reports that typical MoCap jitter can induce inter-frame RTE/RRE up to $\mathcal{F}_{\mathrm{I}}$8 mm and $\mathcal{F}_{\mathrm{I}}$9, while calibration errors can produce ATE/ARE beyond $\mathcal{F}_{\mathrm{W}}$0 mm and $\mathcal{F}_{\mathrm{W}}$1. A common misconception is that raw MoCap is already accurate enough for all benchmark metrics; MoCap2GT rejects that premise and treats GT generation as an estimation problem rather than a direct readout problem.

The core objective is a joint maximum-likelihood estimator over MoCap factors, IMU preintegration factors, and bias random-walk regularizers:

$\mathcal{F}_{\mathrm{W}}$2

$\mathcal{F}_{\mathrm{W}}$3

The IMU term is standard preintegration. Between IMU times $\mathcal{F}_{\mathrm{W}}$4 and $\mathcal{F}_{\mathrm{W}}$5,

$\mathcal{F}_{\mathrm{W}}$6

$\mathcal{F}_{\mathrm{W}}$7

$\mathcal{F}_{\mathrm{W}}$8

The resulting inertial residual is

$\mathcal{F}_{\mathrm{W}}$9

This term suppresses MoCap jitter by forcing short-horizon consistency with high-rate IMU motion increments (Shu et al., 17 Jul 2025).

The complementary MoCap factor couples inertial trajectory, spatial extrinsics, and time alignment:

$\mathcal{F}_{\mathrm{W}}$0

A notable implication is that there is no separate calibration stage in the objective. Calibration emerges because the MoCap residual depends explicitly on $\mathcal{F}_{\mathrm{W}}$1 and $\mathcal{F}_{\mathrm{W}}$2, while the IMU factors constrain the inertial states and gravity alignment (Shu et al., 17 Jul 2025).

3. Continuous-time MoCap modeling and temporal alignment

A distinguishing feature of MoCap2GT is its continuous-time MoCap representation. Rather than linearly interpolating discrete MoCap poses, it uses a cumulative cubic B-spline on $\mathcal{F}_{\mathrm{W}}$3 following Sommer et al. (Shu et al., 17 Jul 2025). For a query time $\mathcal{F}_{\mathrm{W}}$4 with four consecutive control poses,

$\mathcal{F}_{\mathrm{W}}$5

The spline basis is

$\mathcal{F}_{\mathrm{W}}$6

with

$\mathcal{F}_{\mathrm{W}}$7

This yields two benefits emphasized in the paper. The first is a $\mathcal{F}_{\mathrm{W}}$8 smoothness prior that attenuates MoCap jitter while preserving global consistency. The second is better Jacobians of the MoCap residual with respect to time offset than linear interpolation, especially in rotation, which directly improves temporal calibration and ARE/RRE (Shu et al., 17 Jul 2025).

Temporal misalignment is modeled as time-varying rather than constant. Because MoCap and IMU use independent local clocks, the paper argues that clock scale error can cause offset drift exceeding $\mathcal{F}_{\mathrm{W}}$9 ms/minute for consumer-grade IMUs (Shu et al., 17 Jul 2025). MoCap2GT therefore represents the offset as a linear B-spline:

$\mathcal{F}_{\mathrm{M}}$0

The MoCap measurement is queried at

$\mathcal{F}_{\mathrm{M}}$1

This treatment is central to the paper’s claim that GT generation must be temporally consistent with the DUT clock, not merely geometrically plausible in the MoCap frame (Shu et al., 17 Jul 2025).

4. Initialization, robustness, and degeneracy handling

The full batch problem is highly nonconvex, and MoCap2GT places unusual emphasis on initialization (Shu et al., 17 Jul 2025). Existing methods such as Vicon2GT are described as lacking an initialization module and therefore not guaranteeing global convergence. MoCap2GT instead uses a coarse-to-fine initializer.

The initialization begins with coarse constant-offset estimation by angular-velocity correlation, then estimates the extrinsic rotation from a linearized rigid-motion relation:

$\mathcal{F}_{\mathrm{M}}$2

Using left and right quaternion multiplication matrices, this becomes

$\mathcal{F}_{\mathrm{M}}$3

To make this robust, the paper introduces a screw-theory-based kernel

$\mathcal{F}_{\mathrm{M}}$4

with $\mathcal{F}_{\mathrm{M}}$5, and solves the resulting overdetermined homogeneous system by SVD (Shu et al., 17 Jul 2025).

After rotational initialization, velocity, gravity, and extrinsic translation are obtained from linear equations combining MoCap pose relations and preintegration:

$\mathcal{F}_{\mathrm{M}}$6

with closed-form least squares

$\mathcal{F}_{\mathrm{M}}$7

The entire initializer is embedded in RANSAC, and biases are initialized to zero (Shu et al., 17 Jul 2025).

Robustness during nonlinear refinement is further improved by a degeneracy-aware measurement rejection rule. MoCap data are partitioned into time windows of size $\mathcal{F}_{\mathrm{M}}$8, and the maximum rotation angle over each window is computed as

$\mathcal{F}_{\mathrm{M}}$9

If $\mathcal{F}_{\mathrm{G}}$0, the window is declared degenerate. The reported hyperparameters are $\mathcal{F}_{\mathrm{G}}$1 s and $\mathcal{F}_{\mathrm{G}}$2 (Shu et al., 17 Jul 2025). In degenerate windows, the MoCap factor still constrains inertial states but does not contribute gradients to spatiotemporal calibration parameters. This directly addresses low-excitation segments, where translational or rotational observability is weak.

5. Empirical performance and significance for SLAM benchmarking

The empirical claim of MoCap2GT is not merely that IMU fusion reduces jitter, but that joint estimation improves both inter-frame and absolute metrics by estimating spatiotemporal calibration more accurately (Shu et al., 17 Jul 2025). On a real-world benchmark using a high-speed-camera reference system based on global bundle adjustment on a calibration board, the paper reports the following results.

Method	Sufficient motion	Motion degradation
HEC-Filter	ATE 2.251 mm, ARE 0.257°, RTE 0.605 mm, RRE 0.115°	ATE 3.262 mm, ARE 0.291°, RTE 0.610 mm, RRE 0.109°
Vicon2GT	ATE 2.178 mm, ARE 0.242°, RTE 0.217 mm, RRE 0.015°	ATE 3.543 mm, ARE 0.277°, RTE 0.229 mm, RRE 0.017°
Kalibr-M	ATE 1.534 mm, ARE 0.207°, RTE 0.183 mm, RRE 0.013°	ATE 2.135 mm, ARE 0.235°, RTE 0.198 mm, RRE 0.015°
MoCap2GT	ATE 1.466 mm, ARE 0.178°, RTE 0.177 mm, RRE 0.013°	ATE 1.681 mm, ARE 0.173°, RTE 0.182 mm, RRE 0.013°

These numbers support the paper’s main argument that raw MoCap and conventional MoCap–IMU fusion pipelines leave enough residual calibration error to bias ARE and short-horizon relative metrics (Shu et al., 17 Jul 2025). MoCap2GT is reported as the only method that fully meets the target accuracy proposed by the authors using a consumer-grade IMU and MoCap system:

$\mathcal{F}_{\mathrm{G}}$3

The benchmarking consequence is explicit. On TUM-VI, official GT yields large RRE/RTE due to insufficient jitter compensation; on EuRoC, some official sequences suffer calibration degradation producing inflated ATE/ARE (Shu et al., 17 Jul 2025). Recomputing GT with MoCap2GT leads to more plausible and stable SLAM benchmarking results. This suggests that the method improves the fidelity of the reference trajectory rather than the SLAM estimate itself.

Implementation details reported in the paper include optimization in Ceres Solver 2.2, a self-collected setup using a Vicon Viro 2.2 and an ICM-42688-P IMU, accelerometer noise density $\mathcal{F}_{\mathrm{G}}$4, gyroscope noise density $\mathcal{F}_{\mathrm{G}}$5, and MoCap translational/rotational noise densities $\mathcal{F}_{\mathrm{G}}$6 and $\mathcal{F}_{\mathrm{G}}$7 (Shu et al., 17 Jul 2025).

6. Scope, misconceptions, and relation to adjacent MoCap-to-GT work

MoCap2GT belongs to a family of methods that use motion capture to define benchmark-quality reference data, but its target differs sharply from adjacent MoCap-to-GT pipelines. In HUM4D, for example, professional Vicon capture is converted into SMPL-aligned annotations for markerless 4D human motion capture via FBX export, IK-based retargeting in Maya, and strict $\mathcal{F}_{\mathrm{G}}$8 fps to $\mathcal{F}_{\mathrm{G}}$9 fps downsampling; that pipeline is designed for human-body benchmarking, not IMU-centric SLAM GT estimation (Park et al., 14 Apr 2026). SmartMocap, by contrast, jointly estimates human and camera motion from extrinsically uncalibrated RGB cameras in a ground-referenced frame, but it is an RGB markerless mocap system rather than a MoCap–IMU GT estimator (Saini et al., 2022). Mo2Cap2 uses a commercial external multi-view marker-less motion capture system to obtain 3D GT for evaluation of egocentric pose estimation, again serving a different benchmark regime (Xu et al., 2018).

A second misconception arises from name similarity. MoCap2GT is unrelated to arbitrary-skeleton video-to-animation systems such as MoCapAnything V2, which predicts animation-ready local joint rotations for target rigs from monocular video and addresses pose-to-rotation ambiguity using a reference pose-rotation pair (Gong et al., 30 Apr 2026). MoCap2GT is not a retargeting or animation pipeline; it is a benchmark-generation method for SLAM.

The limitations stated for MoCap2GT are equally specific. The method assumes access to raw marker-based optical MoCap poses and synchronized-enough IMU streams from the DUT, plus sufficient rotational excitation for accurate calibration (Shu et al., 17 Jul 2025). Although it handles variable time-offset drift, it still uses a windowed low-order model for that drift. Degenerate motion segments are downweighted for calibration, so performance is best when the dataset contains informative rotation. The authors also note that their own reference evaluation setup is constrained in workspace, and the method is a batch estimator rather than an online system.

A plausible implication is that MoCap2GT is most valuable when benchmark precision is limited by calibration and timestamp consistency rather than by lack of a nominal MoCap reference. In that regime, it reframes “ground truth” from a sensor label into a fused estimate whose accuracy is explicitly optimized for the metrics used to judge modern SLAM systems (Shu et al., 17 Jul 2025).