Mobile Gesture Recognition Eval

---
name: mobile-gesture-recognition-eval
description: This evaluation probes a model's ability to recognize hand gestures from mobile device inertial sensor data (accelerometer and gyroscope). It measures classification robustness across varying signal speeds, amplitudes, and noise levels by testing on three distinct datasets with different gesture sets and sensor configurations. Use when the user wants to benchmark on MGD, BUAA Mobile Gesture Database, SmartWatch Gesture Database, or asks about evaluating this task. Reports classification accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 1707.03692
  bibtex_key: xie2017deepfisher
  confidence: high
---

# mobile-gesture-recognition-eval

> Deep Fisher Discriminant Learning for Mobile Hand Gesture Recognition — Chunyu Xie et al. (2017) (arXiv:1707.03692, 2017)

## What this evaluates

This evaluation probes a model's ability to recognize hand gestures from mobile device inertial sensor data (accelerometer and gyroscope). It measures classification robustness across varying signal speeds, amplitudes, and noise levels by testing on three distinct datasets with different gesture sets and sensor configurations.

## Datasets

- **MGD** — total 5547; splits: train (3500), test (2047)
- **BUAA Mobile Gesture Database** — total 1120; splits: train (784), test (336)
- **SmartWatch Gesture Database** — total 3200; splits: train (2400), test (800)

## Metrics

- `classification accuracy` **(primary)** — range: percent
  - Standard accuracy calculated as the number of correctly predicted gesture sequences divided by the total number of test sequences, reported as a percentage. Results are averaged over 5 independent runs.

## Input / output format

**Input**: 2D tensor of shape (sequence_length, 6) for MGD/BUAA or (sequence_length, 3) for SmartWatch, containing preprocessed (filtered, amplitude-normalized, length-normalized via cubic spline interpolation) accelerometer and gyrometer time-series data.

**Output**: Discrete class label corresponding to one of the predefined gestures (e.g., 'A', 'B', '1', '2', etc.).

## Scoring recipe

```python
def compute_accuracy(predictions, gold_labels):
    correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    return (correct / len(gold_labels)) * 100.0
```

## Common pitfalls

- Sequence lengths vary significantly across datasets after preprocessing (1000 steps for MGD/BUAA vs. 50 steps for SmartWatch), requiring dataset-specific input shaping.
- The SmartWatch dataset only provides 3-axis accelerometer data, whereas MGD and BUAA use 6-axis (accelerometer + gyrometer), necessitating different input channel handling.
- Results on MGD are explicitly averaged over 5 runs, but the paper does not specify the number of runs for the other two datasets, which may affect reproducibility.

## Evidence (verbatim from paper)

> We report the performance of different methods on the testing set based on the average over 5 runs. The class-wise classification accuracy comparison of different mothods is presented in Table.[1] It is clear that by incorporating the Fisher criterion to the base models (BLSTM and BGRU), the recognition performance can be improved.

## Citation

```bibtex
@misc{xie2017deepfisher,
  title={Deep Fisher Discriminant Learning for Mobile Hand Gesture Recognition},
  author={Chunyu Xie et al. (2017)},
  year={2017},
  note={arXiv:1707.03692}
}
```

- arXiv: 1707.03692