RevalExo: A Functional Daily-Activity Benchmark for Inertial and Visual Locomotion Mode Recognition in Older Adults and Clinical Cohorts
Diwas Lamsal, Juha Carlon, Reinhard Claeys, Maxim Yudayev, Louis Flynn, Tom Verstraten, David Beckwée, Eva Swinnen, et al.
cs.AI · cs.CV
Quick verdict
RevalExo claims the first public inertial+visual locomotion benchmark with frame-level labels in older/clinical cohorts; rigorous, but small N and single site.
Summary
RevalExo is a benchmark dataset for locomotion mode recognition combining seven lower-body IMUs and synchronized egocentric video, collected from 27 participants in three cohorts: seven non-impaired older adults, ten stroke survivors, and ten older adults with probable sarcopenia. Recordings follow the FATIG'AGE daily-activity protocol and yield 10.1 hours of frame-level annotations across 11 modes, including 5.1 hours of paired inertial-visual data (video was omitted for 14 participants for clinical safety). Annotation quality is quantified via triple annotation of ~10% of data (Fleiss' kappa = 0.944). The authors benchmark three tasks: (1) recognition with inertial, visual, and fused models under leave-one-subject-out cross-validation, where the best model (MViT-DCL) reaches 92.8% overall macro F1 but only 68.2% during transition windows at tau=0 s; (2) cross-population generalization from healthy older adults to clinical cohorts, where inertial-only models degrade severely (39.9% overall F1 for the sarcopenic cohort) while the multimodal model retains 88.1% on the multimodal stroke subset; (3) vision-guided knowledge transfer to IMU-only models, where contrastive pretraining plus feature distillation adds 7.4 pp (sarcopenia) and 6.9 pp (stroke). Jetson Orin Nano throughput figures assess deployability. Main limitations: small cohorts (N=7-10), a single fixed environment that visual models may partly memorize, and modest, sometimes non-significant transfer gains. Dataset, annotations, and code are released.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- RevalExo is the only public benchmark combining paired inertial-visual recordings, frame-level locomotion labels, and clinical-cohort coverage (stroke survivors and older adults with probable sarcopenia).
- Multimodal inertial-visual fusion yields the best recognition (92.8% overall macro F1 at τ=0 s, MViT-DCL, LOSO-CV) but transition-window performance lags far behind at 68.2%.
- Inertial-only models degrade severely under cross-population transfer (39.9% overall F1 to the sarcopenic cohort at τ=0 s), while visual cues transfer more consistently (88.1% multimodal on the multimodal stroke subset).
- Vision-guided training (contrastive pretraining plus feature distillation) significantly improves IMU-only inference on clinical cohorts, adding 7.4 pp (sarcopenia) and 6.9 pp (stroke) overall F1 at τ=0 s.
Strengths
Unique clinical coverage with paired modalities
Per Table 1 and §2.1, RevalExo is the only public resource combining paired inertial-visual recordings, frame-level labels, and clinical cohorts (stroke FAC 3-4, probable sarcopenia), against prior sets that are healthy-only, IMU-only, or session-labeled (e.g., Sharma et al.).
Quantified annotation reliability
§3.4: triple independent annotation of ~10% of data gives Fleiss' κ = 0.944 (per-class κ > 0.92 except sit-to-stand 0.868, stand-to-sit 0.832), boundary F1 of 0.866 at ±250 ms and 0.957 at ±500 ms; the reliability subset and computation scripts are released.
Deployment-aware evaluation design
Honest confound probing
Inferential testing where it matters most
Weaknesses
Very small cohorts underpin all generalization claims
Single fixed environment
Transitions defined post-hoc, not annotated
Safety-driven selection bias in the multimodal subset
Modest and partially non-significant transfer gains
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
RevalExo (arXiv 2609.08090v2; KU Leuven, VUB, TU Delft) is a benchmark/dataset paper. It releases 10.1 hours of frame-level-annotated daily-activity recordings from 27 participants — 7 non-impaired older adults (HC), 10 stroke survivors (ST), 10 older adults with probable sarcopenia (SR) — performing the FATIG'AGE protocol, with seven lower-body Xsens IMUs (60 Hz) for all and synchronized Pupil Core egocentric video (30 fps, 1280×720) for a 13-subject subset (5.1 h paired). Four claims are central: (1) it is the only public resource combining paired inertial-visual data, frame-level labels, and clinical cohorts (§2.1, Table 1); (2) inertial-visual fusion is strongest, but transitions lag — MViT-DCL (C) reaches 92.8% overall vs 68.2% transition macro F1 at τ=0 s under LOSO-CV (Table 4); (3) inertial-only models degrade sharply under cross-population shift — training on 7 HC yields 57.7% F1 on ST and 39.9% on SR at τ=0 s (Table 6); (4) vision-guided transfer (contrastive pretraining + FitNets distillation) improves IMU-only models on clinical cohorts by +7.4 pp (SR) and +6.9 pp (ST) at τ=0 s (Table 7).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the RevalExo dataset?
RevalExo is a public benchmark of 10.1 hours of frame-level-annotated daily-activity recordings from 27 participants — older adults without mobility impairments, stroke survivors, and older adults with probable sarcopenia — captured with seven lower-body IMUs and, for 13 participants, synchronized egocentric video, covering 11 locomotion modes such as stair ascent, ramps, and uneven ground.
Why is locomotion transition recognition so much harder than steady-state recognition?
Transitions are rare, individually variable, and ambiguous: sensor readings during a switch match neither the previous nor the upcoming mode. In the paper, the best fused model reaches 92.8% overall macro F1 but only 68.2% within ±0.25 s of mode boundaries at zero prediction horizon, underscoring that boundary moments remain an open problem.
Is the RevalExo dataset publicly available, and what is included?
Yes. The paper states that the dataset, frame-level annotations, a triple-annotated reliability subset with computation scripts, and the benchmark codebase are released via revalexo.github.io, and training logs are available on Hugging Face.
How well do IMU-only models work on clinical populations?
Poorly without adaptation. When trained only on seven healthy older adults, an inertial-only DCL model drops to 57.7% macro F1 on stroke survivors and 39.9% on the sarcopenic cohort at zero horizon. Vision-guided pretraining plus distillation improves these IMU-only results by roughly 7 percentage points.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Values embedded in Figures 2, 4, and 5 appear only as interleaved axis/label text in the extraction, so per-model FPS numbers in Figure 5 could not be verified numerically (Appendix B Table 8 FPS values were available and are used instead).
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.