DiffusionRef-IMU: Guide-Anchored Residual Diffusion for Cross-Location Inertial Measurement Unit Translation
Organizations: Department of Computer Science, American International University-Bangladesh (AIUB), Dhaka, Bangladesh · Department of Computer Science and Engineering, Mälardalen University, Västerås, Västmanland, Sweden
Abstract
Cross-location variation in wearable inertial measurement unit (IMU) signals complicates learning when measurements from a desired target body location are unavailable. This study proposes DiffusionRef-IMU, a two-stage cross-location translation framework in which a deterministic Direct network first estimates the target signal and a guide-anchored residual diffusion model refines only the remaining target discrepancy. The evaluation comprises five subject-disjoint activity-location composition-held-out folds, capacity-matched deterministic and flow-based residual baselines, source-only activity conditioning, computational ablation, and true zero-shot evaluation on two independent datasets. DiffusionRef-IMU with 16-step deterministic DDIM achieved a mean nRMSE of 0.087538 +/- 0.004922, compared with 0.090262 for Direct, 0.088847 for deterministic residual refinement, and 0.088417 for eight-step CompFlow, and outperformed each baseline in all five folds. Four-step DDIM retained a mean nRMSE of 0.087820 at 47.62 ms per window, providing a favourable quality-computation trade-off. Replacing ground-truth activity conditioning with a source-only activity predictor (macro-F1 0.8473) increased 16-step DiffusionRef nRMSE only to 0.088190. True zero-shot evaluation produced mixed results: DiffusionRef was marginally favourable on Grouvel2025 but remained worse than Direct on Saldaña2025. Overall, guide-anchored residual diffusion improves controlled cross-location reconstruction while revealing clear limits in zero-shot external transfer.
Figures & tables
| Method | Primary task | Core modelling strategy | Source-to-target conditioning / translation | Evaluation protocol | Relevance / distinction |
| IMUCoCo Zhou et al. [2025] | Flexible-placement pose/HAR | Coordinate-conditioned shared representation from variable on-body IMUs | Placement coordinates condition representation; no paired target-waveform synthesis | Pose estimation and HAR across flexible placements | Strong placement-robust representation model, but not explicit source-to-target signal translation |
| SensorGAN Hussein and Bhat [2024] | Missing wearable-sensor recovery | Conditional GAN reconstructing an unavailable sensor stream | Concurrent observed sensors missing stream | Recovery followed by HAR on public wearable datasets | Explicit sensor reconstruction, but assumes multi-sensor context rather than held activity–location translation |
| SynHAR Uhlenberg et al. [2024] | Synthetic inertial augmentation | Personalised biomechanics and human-surface simulation | Motion/body model virtual IMU; not paired observed-source conditioning | Synthetic/real fusion and HAR transfer | Placement-aware virtual IMU synthesis, but simulation/augmentation rather than learned sensor-to-sensor translation |
| Oppel & Munz Oppel and Munz [2025] | Inertial time-series generation | Denoising diffusion model for multivariate IMU synthesis | Distribution/activity-conditioned generation; no paired source-location target | Synthetic generation plus downstream HAR utility | Recent inertial diffusion generation; different path and reconstruction objective |
| Tong et al. Tong et al. [2024] | Conditional generative transport | Generalised conditional flow matching with flexible source distributions and minibatch OT | Conditioned continuous transport; not wearable-placement specific | Generative-modelling benchmarks | Closest flow formulation, but not Direct-guided cross-location IMU reconstruction |
| Dataset | Subjects / records | Activities | Locations | Native sampling | Evaluation role |
| Comprehensive IMU Feng et al. [2026] | 30 subjects | 12 | 17 | 60 Hz | Five subject-disjoint activity–location composition-held-out folds; 120-sample windows with 50% overlap. |
| Grouvel2025 Grouvel et al. [2025] | 59 subjects / 881 records | 15 available | 9 | 128 Hz | True zero-shot transfer using shared Walk semantics; two-second windows resampled to the main 60-Hz representation. |
| Saldaña2025 Rivas-Caicedo et al. [2025] | 64 retained subjects | 6 available | 3 | 50 Hz | True zero-shot transfer using shared Walk and Bike semantics; two-second windows resampled to the main 60-Hz representation. |
| Variant | Direct guide | Second-stage model | Evaluation role |
| Direct | – | None | Strong deterministic reference and first-stage guide. |
| DetRef | Frozen | One-shot deterministic residual U-Net | Capacity-matched deterministic residual control. |
| CompFlow | Frozen | Conditional residual flow with Heun integration | Generative residual-transport comparator. |
| DiffusionRef-IMU | Frozen | Guide-anchored residual diffusion | Proposed method. |
| Method | F0 | F1 | F2 | F3 | F4 | Mean SD |
| Direct | 0.091271 | 0.082307 | 0.093557 | 0.094887 | 0.089286 | |
| DetRef | 0.090614 | 0.080853 | 0.092351 | 0.093014 | 0.087403 | |
| CompFlow-2 | 0.089580 | 0.080438 | 0.092375 | 0.092621 | 0.086815 | |
| CompFlow-8 | 0.089785 | 0.080433 | 0.092378 | 0.092673 | 0.086814 | |
| DiffusionRef-4 | 0.088920 | 0.080121 | 0.091804 | 0.091926 | 0.086327 | |
| DiffusionRef-16 | 0.088499 | 0.079857 | 0.091766 | 0.091613 | 0.085956 |
| Method | Oracle activity | Predicted activity | nRMSE |
| Direct | 0.090262 | 0.090766 | +0.000504 |
| DetRef | 0.088847 | 0.089477 | +0.000630 |
| CompFlow-8 | 0.088417 | 0.089004 | +0.000587 |
| DiffusionRef-4 | 0.087820 | 0.088498 | +0.000678 |
| DiffusionRef-16 | 0.087538 | 0.088190 | +0.000652 |
| Method | Grouvel2025 | Saldaña2025 |
| Direct | ||
| DetRef | ||
| CompFlow-8 | ||
| DiffusionRef-4 | ||
| DiffusionRef-16 |
Wearable Sensing
Wearable sensing uses body-worn devices to measure motion, physiology, and interaction for health, activity, and assistive applications.
See all 143 papers