IMU-Bench: Benchmarking Wrist IMU Representations Across Kinematic, Semantic and Clinical Tasks
Wrist-worn IMUs power everything from gesture recognition to clinical motor assessment, yet most prior benchmarks evaluate each task in isolation, leaving unclear whether a single pretrained representation can serve all of them. This paper introduces IMU-Bench, a unified evaluation of ten tasks spanning kinematic estimation, semantic retrieval, activity recognition and Parkinson's assessment, and proposes InrtlWrist, a masked autoencoder pretrained on recorded and simulated wrist motion. The core finding is that pretraining yields broadly useful features—leading on nine of ten frozen-probe tasks—but that supervision from one downstream task rarely helps unrelated ones, distinguishing universal representation quality from task compatibility.
The benchmark is well-conceived and the controlled comparison is a genuine contribution, though the empirical gains over training from scratch are uneven and several analytical choices temper the strength of the conclusions. InrtlWrist's dominance under frozen probing is clear: "Among the six matched methods, InrtlWrist achieves the best or joint-best primary score on nine of ten tasks under frozen probing" (Section 5). However, some improvements are marginal—frozen features trail training from scratch on full-body pose and odometry—and the clinical task shows only a modest margin: "its advantage over the frozen random encoder is only 3.5 points, providing more limited evidence of a pretraining benefit" (Section 5.1).
The cross-task transfer analysis is the most valuable component. The paper carefully distinguishes beneficial transfer between kinematic tasks—"full-body supervision, for example, reduces odometry ATE from 0.233 to 0.186 m" (Section 6)—from the failure of supervised sources to approach the pretrained encoder on language retrieval. This separation of representation breadth from task complementarity is conceptually sharp, and the architectural ablation honestly reporting that "Stronger decoders therefore narrow or reverse the measured advantage of pretraining, motivating the light readouts used in the benchmark" (Section 7) is methodologically important.
Several limitations undermine the generality of the findings. The pose and odometry tasks use session splits where participants recur across training and evaluation, so the numbers reflect new recordings of seen people rather than unseen individuals.
The pretraining data audit also notes "including unresolved PRISM–SAMoSA participant overlap" (Section 3.1), while the odometry protocol uses "the true initial orientation and is easier than full inertial navigation" (Appendix B), limiting real-world applicability. Downstream gains are uneven: fine-tuning degrades gesture recognition from 64.7 to 54.2 macro-F1, and joint training improves gesture while harming body pose and retrieval, suggesting that supervision for one target can indeed compromise another.
The evidence supports frozen-probe leadership but paints a more nuanced picture for fine-tuning. InrtlWrist leads "all four activity datasets under frozen probing, with 57.8 mean macro-F1 against 56.2 for LIMU-BERT and 53.3 for training from scratch" (Section 5.1), and retrieval shows a clear pretrained gap over scratch. Yet the full-body pose and odometry benchmarks are not subject-disjoint, and the authors properly caution that "Their inputs and training recipes differ from the matched comparison" (Section 5.2) when citing specialized systems, which blunts the practical argument for replacing task-specific pipelines.
Reproducibility promises are comprehensive but partially deferred. The authors state that "Every number in this paper comes from a registry-driven harness that we release with the benchmark" (Reproducibility statement), and detail architectures and data audits extensively. However, the same statement notes that "per-cell result files and pretrained weights will be released there," implying they were not yet available at submission. Hyperparameters and the 12-layer, 384-wide architecture are fully specified, but independent reproduction currently awaits the promised weights and complete result registry.
Wrist-worn inertial measurement units (IMUs) provide a common sensing modality for tasks ranging from activity and gesture recognition to pose estimation, inertial odometry and clinical motor assessment. Existing wearable benchmarks evaluate performance across multiple tasks, but leave open how supervision for one task affects transfer to others and whether these tasks benefit from a shared representation. We present IMU-Bench, a benchmark of ten tasks from eight datasets for studying IMU multi-task performance and cross-task transfer. We compare six self-supervised methods pretrained on unlabeled recorded and simulated wrist motion, evaluating classification, pose and velocity estimation, and IMU-language retrieval. We find that our masked autoencoder, InrtlWrist, achieves the best or joint-best frozen-probe performance on nine of ten tasks, and fine-tuning improves on training from scratch on nine of ten tasks, including full-body pose, odometry and retrieval. Across transfer and joint-training experiments, supervision that benefits one target can reduce performance on others, with supervised sources transferring within kinematic estimation while, under frozen probing, none approaches the pretrained encoder on language retrieval, and joint-training gains that come at a cost to the paired task. These findings distinguish broad downstream usefulness from complementary task supervision. Code, task cards and a leaderboard are publicly available.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.