Source-linked AI summary
MorphCL: Morphological Contrastive Learning for Inertial-based Human Activity Recognition
Marius Bock, Yuwei Zhang, Juergen Gall, Michael Moeller, Kristof Van Laerhoven, Cecilia Mascolo
TL;DR
Inertial HAR has abundant unlabeled wearable data, but current SSL methods leave global motion structure underused and labeled datasets are costly to build. MorphCL groups unlabeled motion using motifs and domain-specific features, then uses the groups in a contrastive objective. The authors report downstream gains and competitive probing results against a foundation model trained on substantially more data.
Problem
Existing inertial SSL often relies on random batches and local comparisons despite recurring motion structure, while labeled IMU datasets are costly and mostly controlled.
Source: §1 · p. 2Method
MorphCL combines motif discovery and domain-specific feature clustering to define morphology-based positives in a group-based contrastive objective.
Source: §1 · p. 2Results
Up to 15% in F1-score improvement is reported with existing SSL frameworks, and MorphCL encoders beat foundation-model probing results while trained on 4600× less data.
Source: §1 · p. 3, §6 · p. 10Takeaways & Limitations
The results suggest domain-informed inductive bias can complement large-scale training in inertial representation learning.
Source: §1 · p. 3, §6 · p. 10Scalability to substantially larger datasets has not been studied, and the evaluated CAPTURE-24 pretraining scale remains modest relative to private corpora.
Source: §5.5 · p. 9
Abstract
from arXiv · showhide
Despite the ubiquity of sensors in wearable and mobile devices and the abundance of human movement data they generate, translating unlabeled recordings into foundational motion models remains an open challenge. Self-supervised learning (SSL) has alleviated the need for costly annotations, yet existing approaches leave the global structure of large-scale motion data largely untapped, relying on randomly sampled batches and local comparisons that become particularly problematic for in-the-wild inertial data dominated by stationary, low-variance behaviors. Here we introduce Morphological Contrastive Learning (MorphCL), a self-supervised pretraining framework that uses structure-aware grouping to inject explicit modeling of global structure into inertial-based SSL approaches. Building on two well-established pillars of motion analysis, the discovery of motion primitives, or motifs, and domain-specific feature descriptors, we show that MorphCL substantially improves linear probing and finetuning results of learned encoders by up to 15 percentage points in F1-score. In a comparison with existing foundation models, we demonstrate that MorphCL-pretrained encoders match or surpass them models in linear probing performance while trained on $4600\times$ less data. Qualitative analysis of the resulting embedding spaces further reveals morphologically meaningful cluster structure, with improved separation of kinematically similar activity classes.
1 INTRODUCTION
Inertial HAR has abundant unlabeled data but limited labeled, controlled datasets, while common SSL objectives rely on local comparisons that do not reflect recurring motion structure. MorphCL uses motif- and feature-based grouping to define contrastive positives, improving reported downstream results and comparison with a larger-data foundation model.
Motivation
- Retrospective annotation of raw IMU signals is difficult without additional context, so labeled datasets are costly and largely controlled-laboratory recordings.Those datasets lack the scale and diversity needed for modern architectures and real-world behavioral variability.Source: §1 · p. 2
- Instance-level SSL treats each segment independently, although recurring motion primitives are shared across activities and subjects, making that assumption poorly aligned with inertial data.Prior work includes Tang et al. (2021) and Xu et al. (2024).Source: §1 · p. 2
- Randomly sampled batches remain problematic in imbalanced, in-the-wild datasets; about 75% of CAPTURE-24 data corresponds to stationary behaviors.Such batches can contain many near-identical segments, sustaining false negatives even when pairs are chosen carefully.Source: §1 · p. 2
Contribution
- MorphCL discovers motifs and combines them with domain-specific features to group unlabeled motion, then uses those groups to define contrastive positives beyond augmentation pairs.The resulting group-based objective is designed to encourage agreement between segments with similar motion patterns.Source: §1 · p. 2
Results
- Explicitly modeling morphological structure, when combined with existing SSL frameworks, improves downstream results by up to 15% in F1-score and yields representations competitive with a foundation model trained on 4600× more data.Source: §1 · p. 3
2 RELATED WORK
Prior inertial HAR contrastive methods broaden positive-pair selection but remain instance-centric and rely on random batches. MorphCL instead uses domain-informed grouping, while related group-learning approaches and motion foundation models provide distinct points of comparison.
Inertial-based Contrastive Learning
- REBAR uses reconstruction error for semantic similarity and RelCon uses relative ranking, but both remain instance-centric and rely on randomly sampled batches without explicit global structure.REBAR and RelCon (Xu et al., 2024; 2025) expand beyond augmentation-defined positives.Source: §2 · p. 3
Group-based Contrastive Clustering
- MorphCL combines motif discovery with multiple domain features for inertial grouping, extending PPG's sVRI proxy approach because multi-axis inertial dynamics cannot be captured by one scalar proxy.PaPaGei (Pillai et al., 2025) uses sVRI for PPG; MorphCL decouples clustering from representation learning and complements instance-level objectives.Source: §2 · p. 3
Foundation Models for Motion Data
- Wearable foundation-model work includes minute-level handcrafted features or large-scale inertial pretraining; Yuan et al. (2024) is the only model in this set with open-source weights for comparison.SlotFM (Park et al., 2025) uses slot-based attention to decompose motion signals into latent components.Source: §2 · p. 3
3 METHODOLOGY
MorphCL builds morphology-aware pseudo-labels from unlabeled inertial data and uses them in a group-based contrastive objective. Its pipeline discovers repeated patterns, clusters domain-specific features, and encourages agreement among windows assigned to the same morphological group.
3 METHODOLOGY
- MorphCL has three stages: motif discovery, clustering domain-specific features, and a group-based contrastive objective that injects dataset-level structure into existing SSL approaches.The framework constructs morphology-aware pseudo-labels from unlabeled motion and defines positives beyond standard augmentation pairs.Source: §3 · pp. 3–4
3.1 MOTIF-BASED PRE-FILTERING
- Motif discovery uses multidimensional matrix profiling on contiguous chunks to find repeated patterns similar across sensor axes, requiring support from at least two accelerometer axes.A greedy masking strategy suppresses similar reoccurrences of each selected motif before discovering the next, broadening motif diversity.Source: §3.1 · p. 4
3.2 MORPHOLOGICAL CLUSTERING
- Context windows are represented by concatenated statistical activity-recognition features and clustered with HDBSCAN, which determines the cluster count from data density and rejects outliers.Windows assigned to a morphological group provide the group labels used to define positives and negatives in the pretraining corpus.Source: §3.2 · pp. 4–5
3.3 MORPHOLOGICAL CONTRASTIVE LEARNING (MORPHCL)
- MorphCL extends each batch with augmented motif windows and treats windows sharing a morphological label as positives, encouraging agreement at the group level rather than among isolated pairs.The loss is described as a group-based variant of Supervised Contrastive Loss; the selected augmentations preserve the paper's definition of morphology.Source: §3.3 · p. 5
4 EXPERIMENTS
Experiments pair established SSL pretext tasks with MorphCL across two pretraining datasets, then assess pretrained encoders on five downstream HAR benchmarks using linear probing and finetuning. The evaluation uses wrist-sensor data and fold- and seed-averaged LOSO results.
4.1 PRETRAINING
- MorphCL is evaluated with multiple SSL pretext tasks and both DeepConvLSTM and repurposed ViT encoders.The pretext tasks are assessed both with and without MorphCL.Source: §4.1 · p. 5
- WISDM and CAPTURE-24 provide the pretraining corpora, with both resampled to 50 Hz and segmented into non-overlapping 128-sample windows.Each window spans 2.56 seconds; normalization uses the training windows' global mean and standard deviation.Source: §4.1 · pp. 5–6
4.2 DOWNSTREAM TASKS
- Table 1 characterizes downstream datasets by participant and activity counts, average recording duration, and periodic, non-periodic, or complex activity categories.The activity categories distinguish recurring movements, sporadic movements or gestures, and sequential combinations of both.Source: Table 1 · p. 6
4.1 PRETRAINING
- MorphCL forms batches by evenly sampling across K morphological clusters, allowing repeated sampling from smaller clusters to balance cluster representation.MorphCL batches may also be sampled repeatedly so the tested SSL frameworks can sample each training window once per epoch.Source: §4.1 · p. 6
4.2 DOWNSTREAM TASKS
- The evaluation uses five HAR benchmarks, right-wrist sensors resampled to 50 Hz, and retains complex activities and NULL-class background data.The authors include sports and process-related tasks among the more complex recognition settings.Source: §4.2 · p. 6
- Results are fold-averaged Leave-One-Subject-Out cross-validation scores and averaged across three runs.Each downstream participant serves as the test set exactly once.Source: §4.2 · p. 6
5 RESULTS
MorphCL generally improves existing SSL methods, while visualization and foundation-model comparisons indicate more structured embeddings and competitive probing performance. The results also identify limits in the scale and sensor scope evaluated.
5.1 COMPLEMENTARITY TO EXISTING SSL METHODS
- Up to 15% in F1-score: adding MorphCL improved linear-probing results across nearly all evaluated SSL methods.REBAR and RelCon benefited most; MTL improved by as much as 5% on CAPTURE-24, while SimCLR showed no improvement with full-dataset pretraining.Source: §5.1 · p. 7
5.2 ABLATION EXPERIMENTS
- The selected defaults achieved the highest results, and performance remained stable across varied hyperparameter setups; at ρ = 400, all samples became noise and evaluation was prevented.The sensitivity study varied motif-discovery parameters σ and r and clustering parameters ρ and support.Source: §5.2 · p. 7, Figure 2 · p. 8
5.3 QUALITATIVE RESULTS
- On unseen WISDM data, MTL embeddings were entangled, MorphCL formed more coherent manifolds, and their combination produced compact, better-separated clusters.Silhouette coefficients consistently increased with MorphCL; residual errors concentrated among semantically related activities.Source: §5.3 · p. 8, Figure 3 · p. 9
5.4 COMPARISON TO EXISTING ACCELEROMETER FOUNDATION MODELS
- The ViT-S model achieved probing performance comparable to the Biobank model, while the same-backbone ResNet-18 surpassed it despite pretraining on a dataset 4600× smaller.Four consecutive 2.56-second windows were averaged to make embeddings comparable to the foundation model’s 10-second input windows.Source: §5.4 · pp. 8–9, Table 5 · p. 9
5.5 LIMITATIONS
- Scalability to substantially larger datasets remains unstudied, and the framework was evaluated only on accelerometer data.The authors also report that larger ViT variants did not improve linear probing, although finetuning improved for ViT-M.Source: §5.5 · p. 9
6 CONCLUSION
The conclusion presents MorphCL as an auxiliary group-based contrastive loss that complements existing SSL and improves downstream results. The appendices describe motif discovery, morphological clustering, and the pretraining setups used for comparison.
6 CONCLUSION
- Up to 15% in F1-score: adding MorphCL improved representation quality, and the method beat Biobank-model probing results despite using 4600× less data.MorphCL combines motif discovery and domain-specific feature clustering to create morphological groups used by an auxiliary contrastive loss.Source: §6 · pp. 9–10
A.1 MOTIF-DISCOVERY & MORPHOLOGICAL CLUSTERING
- Motif discovery identifies repeated subsequences using matrix profiling, then domain-specific feature vectors are clustered with HDBSCAN to form morphological groups.The procedure uses motif distance threshold σ and extracts 100 occurrences of each of the top 100 motifs per chunk; HDBSCAN treats unassigned windows as noise.Source: §A.1 · p. 14
A.2 PRETRAINING
- Six SSL frameworks were investigated, including MTL, SimCLR, denoising and masked autoencoders, REBAR, and RelCon; MorphCL used balanced batches across morphological groups.Pretraining used batches of 512 windows for most frameworks; REBAR and RelCon used 1,000 epochs to make total windows drawn comparable to other frameworks.Source: §A.2 · pp. 15–16
A.3 DOWNSTREAM EXPERIMENTS
Downstream evaluation uses linear probing and finetuning with participant-held-out cross-validation. The runtime appendix reports the hardware used and the time required for motif discovery.
A.3 DOWNSTREAM EXPERIMENTS
- Linear probing freezes the pretrained backbone and trains a linear classifier, whereas finetuning also adjusts the backbone.Evaluation used Leave-One-Subject-Out cross-validation, 2.56-second windows with 50% overlap, and class-weighted cross-entropy.Source: §A.3 · p. 16
- Each downstream split was trained for 30 epochs, with learning rates reduced when validation loss failed to improve for five epochs.Results were compared with fully supervised training on each dataset and backbone.Source: §A.3 · p. 16
A.4 LICENSING & RUNTIME ENVIRONMENT
- Motif discovery took around 170 seconds per one-hour data chunk on a standard CPU.Experiments used a single NVIDIA Tesla V100 GPU, with larger ViT variants trained on an NVIDIA A6000.Source: §A.4 · p. 16
B.1 DETAILED MAIN RESULTS
The detailed-results tables cover downstream macro-F1 comparisons for MorphCL and instance-based SSL methods across five benchmarks, pretraining corpora, and encoder backbones. Results are averaged across participant-level folds and repeated random seeds.
B.1 DETAILED MAIN RESULTS
- Tables 8 and 9 report relative linear-probing and finetuning improvements from combining MorphCL with instance-based methods.Table 7 reports downstream macro F1 as mean ± standard deviation across seeded runs.Source: §B.1 · p. 17
- Reported scores average macro F1 across Leave-One-Subject-Out splits and three random-seed runs; tables also provide standard deviations across seeds.The last column gives the average across all downstream datasets.Source: §B.1 · p. 17
- Table 7 compares standalone MorphCL with instance-based SSL methods across five downstream benchmark datasets, pretraining on WISDM and CAPTURE-24 with different encoder backbones.The table also includes fully supervised training without pretraining.Source: Table 7 · p. 17
B.2 SENSITIVITY ANALYSIS ON MOTIF DISCOVERY PARAMETERS
Motif discovery and downstream results remain relatively stable across tested motif-distance and occurrence-threshold settings. The selected configuration is σ = 1.0 and r = 100 for motifs of length 0.2 seconds.
B.2 SENSITIVITY ANALYSIS ON MOTIF DISCOVERY PARAMETERS
- MorphCL downstream results remain stable across tested σ and r settings, with σ = 1.0 and r = 100 chosen for 0.2-second motifs.Smaller σ yields fewer motifs and extracted windows; the authors attribute a slight decrease in extracted windows at larger distances to their masking strategy.Source: §B.2 · p. 17
- Motifs longer than 0.2 seconds were not ablated, while motifs shorter than 0.2 seconds were considered too small to capture meaningful motion at 50 Hz.The authors expect WISDM results to represent CAPTURE-24 because motif discovery is applied in one-hour chunks on both datasets.Source: §B.2 · p. 17
- Table 8 reports absolute percentage-point linear-probing improvements from adding a morphological loss across five downstream datasets, two pretraining datasets, and different encoder backbones.Macro F1 is reported as mean ± standard deviation across seeded runs.Source: Table 8 · p. 18
B.3 SENSITIVITY ANALYSIS ON CLUSTERING PARAMETERS
Clustering quality and downstream MorphCL results are stable across moderate settings of ρ and cluster support, but degrade at tested extremes. The authors select ρ = 100 and support = 100.
B.3 SENSITIVITY ANALYSIS ON CLUSTERING PARAMETERS
- DBCV, AMI, and downstream MorphCL results remain stable across ρ and support values from 50 to 200, with ρ = 100 and support = 100 selected.DBCV is a label-free cluster-validity measure; AMI compares clusters with activity labels.Source: §B.3 · p. 18
- At tested values of 400, clustering quality degrades when ρ exceeds WISDM’s density scale or high support merges distinct sub-clusters into coarse groups.The passage reports DBCV and AMI alongside cluster count and noise rate.Source: §B.3 · p. 18
- DBCV evaluates cluster validity without ground-truth labels on a −1 to 1 scale, while AMI measures agreement with activity labels on a 0 to 1 scale.Higher scores indicate better-separated clusters for DBCV and stronger agreement for AMI.Source: §B.3 · p. 18
B.4 SUBSET EVALUATION
Across the reported subset, sampling, and loss-weighting analyses, MorphCL gains persist across pretraining subsets and tested training choices. The CAPTURE-24 comparison separates gains from motif filtering from gains associated with adding the MorphCL objective.
B.4 SUBSET EVALUATION
- On WISDM, filtering alone yields results comparable to full-dataset training, while adding MorphCL improves results across all tested subsets.The authors attribute the gains to the objective rather than subset selection alone.Source: §B.4 · p. 18, Table 14 · p. 20
- All evaluated frameworks benefit from combining with MorphCL in finetuning, with MTL the best-performing method.Table 9 reports seed-averaged absolute finetuning percentage-point improvements across three runs.Source: Table 9 · p. 19
- On CAPTURE-24, motif filtering adds +2.44 for MTL and MorphCL adds a further +2.89 on the same subset, totaling +5.33 over the full-corpus baseline.The passage says filtering and the MorphCL loss each contribute roughly half the total gain.Source: §B.4 · pp. 19–20, Table 15 · p. 20
B.5 MORPHCL SAMPLING METHOD
- MorphCL remains robust to different balancing strategies on WISDM and CAPTURE-24, including training without oversampling or balanced batches.SimCLR slightly benefits from unbalanced MorphCL batches; the authors hypothesize this relates to less gradient collision between losses.Source: §B.5 · p. 20, Table 16 · p. 21, Table 17 · p. 21
B.6 LOSS TERM WEIGHTING
- Adding MorphCL improves results for λ > 0, while unweighted summation (λ = 1) shows no significant difference from weighted sums (λ > 1).The reported comparison uses a ViT-S backbone with MTL; the authors conclude results are robust to the MorphCL loss scaling factor.Source: §B.6 · pp. 20–21, Table 18 · p. 21
B.7 EXTENDED QUALITATIVE ANALYSIS
Extended qualitative analyses report clearer embedding separation and reduced confusion in RWHAR and Hang-Time when MorphCL is combined with MTL. The loss comparison finds the proposed group-based formulation has the highest average F1 in most downstream experiments.
B.7 EXTENDED QUALITATIVE ANALYSIS
- Combining MorphCL with MTL improves separation of kinematically close classes in RWHAR and Hang-Time and reduces confusion in linear-probing results.The plots include UMAP features, silhouette coefficients, and confusion matrices; the table reports increased mean silhouette coefficients across datasets for each evaluated configuration with MorphCL.Source: Figure 4 · p. 21, §B.7 · p. 21, Figure 5 · p. 22
B.8 SUPCON VS. GROUP-BASED AVERAGING
- The group-based MorphCL loss yields the highest average F1 across downstream experiments with few exceptions, while SupCon and Pillai et al. (2025) compute losses over individual positive pairs.The alternative loss formulations normalize against the full batch; Table 20 compares them within the MorphCL framework.Source: §B.8 · pp. 21, 23, Table 20 · p. 24
- The comparison evaluates loss formulations using MorphCL alone and in combination with the SSL frameworks reported in the main paper.Results are averaged across three runs.Source: §B.8 · p. 23, Table 20 · p. 24