When the target body location has no sensor, DiffusionRef-IMU lets diffusion fix only the residual and beats deterministic baselines in all five folds
Synopsis
Moving a wearable inertial sensor to another body location shifts the signal distribution, and measurements at the desired target location are often unavailable during training or deployment. DiffusionRef-IMU first has a deterministic Direct network estimate the target signal, then lets a diffusion model correct only the residual between that estimate and the paired target. Across five subject-disjoint folds with activity–location compositions held out, 16-step DDIM reached a mean nRMSE of 0.087538, below 0.090262 for Direct, 0.088847 for a capacity-matched deterministic residual model and 0.088417 for eight-step CompFlow, and was better in all five folds.
Interpretation
Cross-location inertial translation no longer asks a diffusion model to generate the whole target waveform from scratch; the generative stage is confined to the part of the discrepancy the deterministic estimate leaves unexplained. Earlier virtual-sensing and generative inertial work mainly targets data augmentation, missing-sensor recovery or placement-invariant classification, not reconstruction of the specific target-location waveform paired with a given source-location window. Across five subject-disjoint folds with activity–location compositions held out, 16-step DiffusionRef reached a mean nRMSE of 0.087538, against 0.090262 for Direct, 0.088847 for the capacity-matched deterministic residual model DetRef and 0.088417 for eight-step CompFlow, and was better than each in 5/5 folds.
The two-stage pipeline first has a one-dimensional U-Net regress a six-channel target estimate from the source window, then a residual diffusion network sees the source window, that frozen estimate and the current noisy residual at every step. The residual network uses the same parameterisation as the control experiments, with an 18-channel input of six noisy residual channels, six source channels and six Direct-guide channels, and 5,991,046 trainable parameters. Inference uses zero-prior deterministic DDIM; on a Tesla T4 the four-step setting reached 0.087820 mean nRMSE at 47.62 ms per window, 16 steps reached 0.087538 at 161.11 ms, and 32 steps only reached 0.087534 while latency rose to 316.15 ms.
Perspective
The demonstrated scope is limited to six-channel accelerometer–gyroscope signals, the 12 studied activities and 17 body locations, two-second windows and the Comprehensive IMU dataset participants; external checks cover only Grouvel2025 and Saldaña2025 and use only activity semantics present in the main vocabulary.
The external zero-shot evidence is itself mixed: DiffusionRef-16 is marginally better than Direct on Grouvel2025 but worse on Saldaña2025, and frozen main-data normalisation clips roughly 0.55 and 0.48 of samples on the two external sets, indicating substantial distribution shift. With five paired folds the exact two-sided sign-flip value bottoms out at 0.0625, so direction, effect size and confidence intervals carry more information than a significance threshold.