Research map / Architectures and efficiency / RNNs and normalization
HEROS-GAN: Honed-Energy Regularized and Optimal Supervised GAN for Enhancing Accuracy and Range of Low-Cost Accelerometers
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 HEROS-GAN, a CycleGAN-based unpaired GAN with optimal transport supervision and a modulated Laplace energy regularizer, to enhance accuracy and extend range of low-cost smartphone accelerometers, and releases the LASED dataset.
Problem
Low-cost accelerometers suffer from high noise and limited measurement range (saturation at ±2g/±8g), and paired low-cost/high-cost frame-level training data is unavailable.
Approach
Combines Sinkhorn-based optimal transport alignment of unpaired low-cost and high-cost signal features (OTS) with a Laplace-energy regularizer whose modulation parameter is set to feature kurtosis (MLE) inside a CycleGAN generator to inject appropriate local variation and break range limits.
Key results
- HEROS-GAN achieves CSRE at tau=6g of 0.329 vs 1.519 for best prior (CNN-o), a 77% reduction relative to best baseline.
- HEROS-GAN reduces ZVRE to 21.05 m/s from 350.78 raw, a 94.0% reduction.
- Allan variance BI reduced to 0.065 from raw 3.119 (-97.9%), versus -85.8% for the unsupervised IMUDB baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HEROS-GAN instead of the strongest listed baseline | LASED clipped signals at τ=6g · LASED, clipped signals (τ=6g) | CSREτ (g): 1.519 -> 0.329 | Helped |
| Use HEROS-GAN instead of the strongest listed baseline | LASED saturated low-cost signals · LASED, saturated low-cost signals | ZVRE (m/s): 310.4 -> 21.05 | Helped |
| Use HEROS-GAN instead of IMUDB | LASED static signals · LASED, static signals | QN: 0.189 -> 0.057 | Helped |
| Use HEROS-GAN instead of IMUDB | LASED static signals · LASED, static signals | VRW: 0.274 -> 0.06 | Helped |
| Use HEROS-GAN instead of IMUDB | LASED static signals · LASED, static signals | BI: 0.444 -> 0.065 | Helped |
| add MLE to CycleGAN (ablation) | CycleGAN · LASED, clipped signals (τ=6g) | CSREτ (g): 1.803 -> 0.474 | Helped |
| add MLE to CycleGAN (ablation) | CycleGAN · LASED, saturated low-cost signals | ZVRE (m/s): 332.8 -> 195.7 | Helped |
| add OTS to CycleGAN (ablation) | CycleGAN · LASED, saturated low-cost signals | ZVRE (m/s): 332.8 -> 39.41 | Helped |
| replace OTS with L1 supervision (ablation) | CycleGAN · LASED, clipped signals (τ=6g) | CSREτ (g): 1.803 -> 2.091 | Hurt |
| replace OTS with L1 supervision (ablation) | CycleGAN · LASED, saturated low-cost signals | ZVRE (m/s): 332.8 -> 433.9 | Hurt |
| add MLE and OTS (ablation) | CycleGAN · LASED, clipped signals (τ=6g) | CSREτ (g): 1.803 -> 0.329 | Helped |
| add MLE and OTS (ablation) | CycleGAN · LASED, saturated low-cost signals | ZVRE (m/s): 332.8 -> 21.05 | Helped |
Surprising findings
- Fully supervised deep models trained on clipped/unclipped high-cost pairs degrade static noise metrics (e.g., CNN-o BI increased 10.2% over raw signal).
Limitations
- Training uses data from only one smartphone type; generalization is tested on nine other phones.
- Evaluation relies on self-defined metrics (CSRE, ZVRE) and clipping simulation of high-cost signals.
- Claimed theoretical convergence and GAN stability proofs are only referenced in the conclusion, not detailed in the main text.
Experimental setup
Datasets: LASED
Benchmarks: LASED
Metrics: CSRE ZVRE Allan variance (QN, VRW, BI)
Baselines: EMD-Kalman filter Savitzky-Golay filter Matlab 2023 over-range reconstruction CNN-o GRU-LSTM GRU-LSTM-o kNN IMUDB CycleGAN
Builds on: CycleGAN Sinkhorn algorithm optimal transport
Closest papers by meaning
- Wavelet Dynamic Selection Network for Inertial Sensor Signal Enhancement (AAAI 2024 · similarity 0.80)
- MCGAN: Enhancing GAN Training with Regression-Based Generator Loss (AAAI 2025 · similarity 0.80)
- Inertia-1: An Open Exploration of Wearable Motion Foundation Models (NeurIPS 2026 · similarity 0.80)
- UniMTS: Unified Pre-training for Motion Time Series (NeurIPS 2024 · similarity 0.80)
- Scaling Wearable Foundation Models (ICLR 2025 · similarity 0.79)
- Two-stream Beats One-stream: Asymmetric Siamese Network for Efficient Visual Tracking (AAAI 2025 · similarity 0.79)
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.