Skip to content

Real inference on the visitor's device

From wrist IMU signals to activity records

Run the bundled 120-second synthetic example, or replace it with a compatible 100 Hz wrist-IMU recording. Nothing to install, and no sensor data is uploaded.

Real public models WebGPU / WASM Timeline + CSV Data stays local
Run the Demo Browser source Model weights
Activity likelihood curves and final activity timeline for the bundled synthetic example Bundled-example reference output

Run it here

The bundled example is ready. Keep the defaults and choose Run current recording to execute the complete three-scale pipeline on this device.

1 · Recording
2 · Pipeline settings

Local by design. The page downloads public, checksum-verified model files from Hugging Face. Your IMU samples are parsed, filtered, inferred, plotted, and exported inside a Web Worker on this device.

Complete an inference run in three steps

1

Confirm the bundled example

synthetic_activity_imu.tsv is already loaded with 12,000 samples and 120 seconds of six-channel IMU data.

2

Keep the defaults

Start with local-boundary fusion, a 5-second minimum duration, 0.30 confidence, and Top-K 5.

3

Run and inspect

Choose Run current recording, then inspect raw signals, the likelihood timeline, and activity records.

Input filesynthetic_activity_imu.tsv
Fusionlocal_boundary
Minimum duration5 s
Confidence0.30
Top-K5

After you choose Run

The Worker validates and zero-phase Butterworth-filters the recording, evaluates the real 3-, 5-, and 8-second ONNX models, then performs scale fusion, Viterbi decoding, boundary refinement, and record filtering.

Bundled-example reference output

With the defaults above, the public models produce 118 timeline points and two activity records. Use these values to confirm that the browser pipeline is operating correctly.

12,000input samples
120.0 srecording duration
118timeline points
2activity records
Activity Start (s) End (s) Duration (s) Confidence
Fly 29.84 73.15 43.31 0.4038
Running 76.06 98.24 22.18 0.3186

A pipeline demonstration, not an accuracy test

The bundled file is deterministically generated and contains no participant data. Its predicted classes and boundaries are model outputs, not ground-truth labels or accuracy evidence.

Read the three result tabs

Raw signals

Start with the three accelerometer and three gyroscope channels to check completeness, sampling stability, and visible movement.

Likelihood + timeline

The upper chart shows six smoothed class probabilities; the lower stepped trace shows the final class index after multi-scale fusion and Viterbi decoding.

Activity records

Each row reports the activity, start, end, duration, and confidence and can be downloaded directly as CSV.

How parameters change the result

Parameter Default Effect
Multi-scale fusion local_boundary Combines the 3-, 5-, and 8-second posteriors, with the largest effect near activity transitions.
Minimum duration 5 s Removes shorter intervals; larger values favor longer activities.
Confidence 0.30 Removes weaker records; raising it may reduce both false positives and recall.
Top-K 5 Limits the number of returned records; 0 keeps all records.

These are Demo settings chosen to make the 120-second example easy to inspect, not immutable activity definitions from the paper experiments.

Upload your own recording

  1. Prepare a UTF-8, tab-separated .txt or .tsv file.
  2. Choose or drop the file and keep the defaults for the first run.
  3. Inspect Raw signals before interpreting model output; change only one parameter at a time afterward.

Required column names:

ACC_TIME  ACC_X  ACC_Y  ACC_Z  GYRO_X  GYRO_Y  GYRO_Z

ACC_TIME must contain strictly increasing millisecond timestamps with a median interval of 8–12 ms. Extra columns and invalid rows are ignored. The public Demo accepts 800–60,000 valid samples (about 8 seconds to 10 minutes at 100 Hz).

Local compute and privacy

What the browser downloads

The first run downloads about 17 MB of ONNX weights and a roughly 26 MB runtime. Files are SHA-256 verified and cached when browser storage is available.

Where the data goes

Parsing, filtering, inference, plotting, and CSV export all run in a Web Worker on this device; the IMU recording is never uploaded.

Models are pinned to Hugging Face revision e0f89bb6…, with checksums published in model-assets.json. WebGPU is preferred when available; otherwise the Demo falls back to WASM CPU, so runtime varies across computers and phones.

Research output

A valid file format does not guarantee compatibility with the training protocol. Sensor placement, axis orientation, units, device characteristics, and preprocessing must match. Predictions are not medical, safety, or coaching advice.

For a server-side comparison, use the original Hugging Face Space. This page uses visitor compute and keeps the recording local.