Research map / Label-efficient and robust learning / Domain adaptation and generalization
SuDA: Support-based Domain Adaptation for Sim2Real Hinge Joint Tracking with Flexible Sensors
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SuDA, a support-based domain adaptation method that registers the supports of simulated and real predictive functions (via quantized curve parameterization) to enable Sim2Real hinge joint angle tracking with flexible sensors without any real labels.
Problem
Flexible-sensor human motion capture needs large labeled real datasets that are costly to collect; distribution-based domain adaptation fails for Sim2Real transfer with low-dimensional sensor data.
Approach
Instead of aligning instance-dependent feature distributions (DiDA), align the supports of the source and target predictive functions by parameterizing each support curve with a 1-D parameter and registering points with equal parameter values, which gives the same joint-angle label.
Key results
- SuDA achieves 7.60 ± 2.58° MAE versus 6.42 ± 2.73° for supervised learning and 21.93° for source-only on the real dataset.
- SuDA outperforms DAN, DANN, D-CORAL, DAAN and BNM, which stay near source-only (about 21.5–22.4° MAE).
- In real-world applications, SuDA reaches 10.74° average MAE versus 12.05° for supervised learning and about 20–21° for DiDA baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use support-based adaptation from simulated data instead of real data from a random user, motion, and wearing position | SuDA hinge-joint tracking · Real dataset test set | MAE (deg): 11.91 ± 4.87 -> 7.60 ± 2.58 | Helped |
| Use SuDA instead of supervised learning | Hinge-joint tracking · Real dataset test set | MAE (deg): 6.42 ± 2.73 -> 7.60 ± 2.58 | Hurt |
| Use support-based domain adaptation instead of source-only training | Hinge-joint tracking · Real dataset test set | MAE (deg): 21.93 ± 8.83 -> 7.60 ± 2.58 | Helped |
| Use SuDA instead of DAN | Hinge-joint tracking · Real dataset test set | MAE (deg): 22.38 ± 8.34 -> 7.60 ± 2.58 | Helped |
| Use SuDA instead of DANN | Hinge-joint tracking · Real dataset test set | MAE (deg): 21.51 ± 8.03 -> 7.60 ± 2.58 | Helped |
| Use SuDA instead of D-CORAL | Hinge-joint tracking · Real dataset test set | MAE (deg): 21.89 ± 8.85 -> 7.60 ± 2.58 | Helped |
| Use SuDA instead of DAAN | Hinge-joint tracking · Real dataset test set | MAE (deg): 21.80 ± 8.28 -> 7.60 ± 2.58 | Helped |
| Use SuDA instead of BNM | Hinge-joint tracking · Real dataset test set | MAE (deg): 21.70 ± 8.61 -> 7.60 ± 2.58 | Helped |
| Apply SuDA to new participants and real-world activities | Hinge-joint tracking · Real-world activities | MAE (deg): 20.61±0.90 -> 10.74±0.67 | Helped |
| Apply SuDA using right-elbow simulation data to other joints | Hinge-joint tracking · Four joints | MAE (deg) | Helped |
Surprising findings
- On several subsets SuDA beats the supervised baseline (e.g., Table 5 for Run and Walk-free motions, and Table 6 average), though the paper attributes this to fluctuation rather than a claim of superiority.
- Distribution-based DA methods perform no better than source-only, and their transfer loss does not decrease during training (Fig. 6).
Limitations
- Assumes source and target function supports have similar ranges of joint angles; fails in extreme cases where effective angle ranges differ.
- Only evaluated on low-dimensional (two-sensor) hinge joint data; generality to high-dimensional data is untested.
- Real dataset is private and the approach requires calibration.
Experimental setup
Datasets: Mixamo motion capture (simulated via SMPL) in-house real flexible sensor dataset
Benchmarks: in-house real flexible sensor dataset (11 users, 12 wearing positions, 4 motions) real-world applications (table tennis, basketball, boxing) multi-joint application dataset
Metrics: MAE (deg)
Baselines: Source Only DAN DANN D-CORAL DAAN BNM Supervised learning Real2Real
Builds on: Deep Domain Confusion Deep Adaptation Networks Deep CORAL DANN SMPL
Closest papers by meaning
- Fast Adaptation for Human Pose Estimation via Meta-Optimization (CVPR 2024 · similarity 0.83)
- MotionPRO: Exploring the Role of Pressure in Human MoCap and Beyond (CVPR 2025 · similarity 0.82)
- Optimal-state Dynamics Estimation for Physics-based Human Motion Capture from Videos (NeurIPS 2024 · similarity 0.82)
- MODA: Motion-Drift Augmentation for Inertial Human Motion Analysis (CVPR 2025 · similarity 0.82)
- MMVP: A Multimodal MoCap Dataset with Vision and Pressure Sensors (CVPR 2024 · similarity 0.82)
- JODA: Composable Joint Dynamics for Articulated Objects (NeurIPS 2026 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.