tracker-pov

Eidon Tracker POV

1,274 hours of egocentric video paired with 7-point IMU arm tracking, recorded during ordinary household work.

Contributors wore a head-mounted camera and a seven-sensor IMU harness while doing real chores in their own homes: laundry, cleaning, dishes, cooking. Each recording pairs first-person video with 24 Hz orientation data for both hands, both forearms, both upper arms, and the chest.

This is a complete, final release. Eidon AI (Solidic Labs Inc) has wound down; the dataset is published under CC-BY-4.0 so it outlives the company.

This repo holds thevideo. The sensor data is next door.

The release is split across three artifacts:

	Contents	Size
this repo (tracker-pov)	the 13,451 MP4s + metadata.parquet	9.05 TB
tracker-pov-imu	the IMU streams for these same recordings, 779M rows	9.5 GB
egocentric-pov (bucket)	extra video with no sensor data	~1.75 TB

The first two are one dataset in two pieces, covering the same 13,451 recordings and joined on recording_id. If you want paired video + motion, you want both. See the organization page for the overview.

Structure

The release is split across two repos and a bucket.

├── README.md
├── recordings/
│   ├── metadata.parquet        # one row per recording
│   └── data/000 … 013/         # 13,451 MP4s, 1,000 per directory

The data/NNN/ split exists only to keep directories manageable; the grouping carries no meaning.

from datasets import load_dataset

recs = load_dataset("eidon-ai/tracker-pov", split="train")
imu  = load_dataset("eidon-ai/tracker-pov-imu", split="train", streaming=True)

Both are keyed by recording_id. Filter the metadata first, then pull only the IMU you need:

import pandas as pd
meta = pd.read_parquet("hf://datasets/eidon-ai/tracker-pov/recordings/metadata.parquet")
cooking = meta[(meta.task_type == "cooking") & (meta.qc_status == "valid")]

imu = pd.read_parquet("hf://datasets/eidon-ai/tracker-pov-imu/imu-0000.parquet",
                      filters=[("recording_id", "in", cooking.recording_id.tolist())])

metadata.parquet

Sensor data

The IMU streams live in eidon-ai/tracker-pov-imu It holds one row per (recording, timestamp, body slot), around 780M rows, joined to this repo on recording_id. Full schema is documented on that card.

Body slots, in fixed order:

The chest sensor is the natural reference frame: composing chest⁻¹ · limb gives torso-relative arm pose, which is invariant to which way the wearer is facing.

Video and IMU synchronisation

Camera and IMU capture were started back to back in the same app, with no hardware sync. The offset is a few milliseconds of sequential code execution. Align by timestamp:

frame N  ->  t_ms = (N / video_fps) * 1000
          ->  nearest row = argmin |time_ms - t_ms|

Activity types

Quality control

Every recording was scored by an automated pipeline before inclusion. Nothing unreviewed is present.

invalid recordings are included deliberately. They are labelled, and negative examples are useful. Filter on qc_status for a clean training set.

Limitations

Read this section before training on the data.

The task distribution is severely skewed. Folding laundry is 67.5% of all hours. Treating this as a balanced nine-class activity dataset will not work; the tail classes have under three hours each.

Contributors are few and unevenly weighted. 27 people. The largest contributes 16.1% of all hours, the top five contribute 57.2%. Split by contributor_id, not randomly. A random split leaks the same person, home, and camera rig across train and test, and will flatter your numbers.

Demographics are sparse and unrepresentative. 16 of 27 reported age (18 to 63, median 26); 17 reported gender, of whom 15 are male. Do not treat this as a population sample.

Raw motion data covers a minority of recordings. Accelerometer, gyroscope and magnetometer readings track a per-contributor opt-in, not sensor_schema_version. No v1 recording has them, and only some v2 recordings do. Filter on has_raw_motion in metadata.parquet, or check for nulls in the imu subset; do not infer it from the version field. Orientation quaternions are present throughout.

A small number of recordings have an incomplete rig. Roughly 0.5% stream fewer than seven slots, sometimes missing the chest sensor that torso-relative pose depends on. Filter on n_slots == 7 and has_chest if you need the full rig.

Some devices are hub-relayed. Sensors may connect through another tracker acting as a BLE hub rather than directly to the phone. The relayed streams are independently measured, not interpolated or duplicated. This is a connection-topology detail, not a data-quality one.

A .json.gz sidecar format was considered and dropped. Earlier drafts shipped per-recording sensor JSON. Parquet is both smaller and actually loadable, so the sidecars do not exist. If you find a reference to them anywhere, it is stale.

Video is unprocessed. See below.

Companion: video-only footage

A further ~324 hours of QC'd egocentric video without sensor data is published as a Storage Bucket at eidon-ai/egocentric-pov (~1.75 TB). Same contributors, same activities, same QC pipeline, just no IMU harness.

It is a bucket rather than a dataset repo because it is unpaired bulk footage with thin metadata. That means no dataset viewer, no load_dataset, and no version history. Fetch it with hf buckets sync or the S3-compatible API. If you want the paired data, you want this repo, not that bucket.

Provenance and consent

Contributors were adult volunteers who gave written, informed consent covering research, commercial use, and redistribution in public datasets, with explicit acknowledgement that egocentric video may constitute biometric data under BIPA and GDPR (processed on the basis of GDPR Art. 9(2)(a) explicit consent). Only recordings from contributors who affirmatively accepted those terms are included.

Personal identifiers have been removed from the metadata and the sensor streams: contributors appear only as integer ids, device MAC addresses and connection names have been replaced by body-slot names, and absolute timestamps have been converted to offsets from the start of each recording.

The video itself has not been altered. It was recorded in private homes, and although contributors were instructed not to record other people, incidental capture of faces, documents, screens, or interiors cannot be ruled out. No face blurring or region redaction has been applied. Anyone deploying models trained on this data should weigh that.

To request removal of specific recordings, contact [email protected].

Citation

@misc{eidon2026trackerpov,
  title  = {Eidon Tracker POV: Egocentric Video with Seven-Point IMU Arm Tracking},
  author = {{Solidic Labs Inc (Eidon AI)}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/eidon-ai/tracker-pov}
}

Licence

CC-BY-4.0. Use it for anything, including commercially. Attribution required.