README.md

---
annotations_creators: []
language: en
license: cc-by-4.0
size_categories:
- n<1K
task_categories:
- robotics
task_ids: []
pretty_name: fomo-multimodal-sample
tags:
- autonomous-navigation
- fiftyone
- lidar
- mcap
- multi-season
- multimodal
- odometry
- off-road
- radar
- robotics
- slam
description: A 6-episode multimodal (MCAP) sample built from the FoMo dataset (Boxan
  et al. 2026, arXiv:2603.08433), a year-long multi-season robot navigation dataset
  recorded in a boreal forest near Quebec City. Each sample is a 30-second trimmed
  episode authored as a single .mcap file combining stereo + mono camera, dual lidar,
  FMCW radar, dual IMU, audio, wheel odometry, and PPK-GNSS ground-truth pose, one
  per trajectory type across 6 distinct deployments spanning Nov 2024-Sep 2025.
dataset_summary: '




  This is a [FiftyOne](https://github.com/voxel51/fiftyone) dataset with 6 samples.


  ## Installation


  If you haven''t already, install FiftyOne:


  ```bash

  pip install -U fiftyone

  ```


  ## Usage


  ```python

  import fiftyone as fo

  from fiftyone.utils.huggingface import load_from_hub


  # Load the dataset

  # Note: other available arguments include ''max_samples'', etc

  dataset = load_from_hub("Voxel51/fomo-multimodal-sample")


  # Launch the App

  session = fo.launch_app(dataset)

  ```

  '
---

# Dataset Card for fomo-multimodal-sample

![image/png](fomo_dataset.gif)

This is a [FiftyOne](https://github.com/voxel51/fiftyone) dataset with 6 samples.

## Installation

If you haven't already, install FiftyOne:

```bash
pip install -U fiftyone
```

## Usage

```python
import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/fomo-multimodal-sample")

# Launch the App
session = fo.launch_app(dataset)
```

## Dataset Details

### Dataset Description

This dataset is a curated, budget-capped **FiftyOne multimodal sample** built from
the [FoMo dataset](https://fomo.norlab.ulaval.ca) (Boxan et al. 2026), a year-long
robotic data collection recorded in Forêt Montmorency, a boreal forest 80 km north
of Quebec City, Canada. The original FoMo dataset totals **9.43 TB across 60
sessions** (12 seasonal deployments × up to 6 repeated trajectories); this sample
selects **one 30-second episode per trajectory type across 6 distinct deployments**
(Nov 2024 – Sep 2025, spanning winter/-19°C through summer conditions), each
authored as a single `.mcap` file combining every raw sensor stream the original
session recorded: stereo + mono camera, two lidars, FMCW radar, two IMUs, stereo
audio, wheel odometry, and PPK-GNSS ground-truth pose. Total sample size is ~19.8 GB
(6 episodes), selected to fit a 30 GB recon budget while maximizing both trajectory
and seasonal diversity.

- **Curated by:** harpreetsahota (this FiftyOne multimodal conversion and 6-episode
  sample selection); original sensor data collected by Norlab (Université Laval)
  and the University of Toronto Robotics Institute
- **Funded by:** Natural Sciences and Engineering Research Council of Canada
  (NSERC) and Fonds de recherche du Québec (FRQNT) grant 2023-NOVA-326877 (HUNTER);
  Canada Foundation for Innovation Fund grant #39709 (PI: E. Thiffault and F.
  Anctil); weather data from the Adaptable Earth Observation System project
  (Canada Foundation for Innovation, Government of Quebec, McGill, UQAM)
- **Shared by:** harpreetsahota
- **Language(s):** N/A (sensor/robotics telemetry, no natural language content)
- **License:** CC BY 4.0

### Dataset Sources

- **Repository:** [s3://fomo-dataset](https://registry.opendata.aws/fomo-dataset)
  (original raw data, public/unsigned S3 access); devkit at
  [fomo.norlab.ulaval.ca](https://fomo.norlab.ulaval.ca)
- **Paper:** Boxan et al. (2026). *FoMo: A Multi-Season Dataset for Robot
  Navigation in Forêt Montmorency.* arXiv:[2603.08433](https://arxiv.org/abs/2603.08433)
- **Demo:** [fomo.norlab.ulaval.ca](https://fomo.norlab.ulaval.ca)

## Uses

### Direct Use

Multi-sensor visualization and exploration in FiftyOne's multimodal viewer
(Image, 3D, Map, Plot, Logs, and Message tiles) — inspecting synchronized
camera/lidar/radar/IMU/pose data for a single robot-navigation episode without
downloading the full 9.4 TB source dataset; prototyping or testing multimodal
`.mcap` ingestion pipelines for off-road robot navigation data; qualitative
comparison of seasonal appearance changes (snow depth, lighting, vegetation)
across the 6 included trajectory/deployment pairs.

### Out-of-Scope Use

Not suitable for training object detection, segmentation, or classification
models — the source FoMo dataset carries **zero** object/semantic annotations
(confirmed by inspecting the full 60-session bucket structure and by keyword
search of the source paper text). Not representative of the full dataset's scale
or statistical diversity for SLAM/odometry benchmarking — this is a 6-episode,
30-second-per-episode recon sample, not the complete 60-session, multi-minute-per-
session dataset; for actual benchmarking, use the full source S3 bucket.

## Dataset Structure

This is a **multimodal** FiftyOne dataset (`dataset.media_type == "multimodal"`)
with **6 samples**, each an authored `.mcap` episode file. There are no
train/val/test splits — every sample is an independent 30-second episode.

### Sample fields

| Field | FiftyOne type | Description |
|-------|---------------|-------------|
| `filepath` | `StringField` | Path to the sample's `.mcap` episode file |
| `trajectory` | `StringField` | Trajectory name (`red`, `blue`, `green`, `magenta`, `yellow`, `orange`) |
| `trajectory_description` | `StringField` | Human-readable terrain/length description of the trajectory |
| `deployment_date` | `StringField` | Source deployment folder date, `YYYY-MM-DD` |
| `session_name` | `StringField` | Original FoMo session folder name, `<trajectory>_<YYYY-MM-DD-HH-mm>` |
| `episode_name` | `StringField` | Local episode identifier used for this sample |
| `window_s` | `IntField` | Trimmed episode duration in seconds (always 30) |
| `anchor_unix_us` | `IntField` | Episode start time, UNIX microseconds (anchored to the first real lidar frame, not the GT log start, which can lag/lead by tens of seconds) |
| `gt_origin_utm` | `ListField` | `[x, y, z]` absolute UTM-like coordinates of the episode's first ground-truth pose, subtracted out before logging `/gt_pose` so positions stay small and float-precision-safe |
| `mcap_bytes` | `IntField` | Size of the episode's `.mcap` file in bytes |
| `n_static_transforms` | `IntField` | Number of static sensor-extrinsic transforms logged to `/tf_static` (always 9) |
| `n_basler_frames`, `n_zedx_left_frames`, `n_zedx_right_frames` | `IntField` | Frame counts for the mono (Basler) and stereo (ZED X) cameras |
| `n_robosense_scans`, `n_leishen_scans` | `IntField` | Scan counts for the two lidars |
| `n_navtech_frames` | `IntField` | Radar scan count |
| `n_audio_left_clips`, `n_audio_right_clips` | `IntField` | 1-second audio clip counts per microphone |
| `n_vectornav_samples`, `n_xsens_samples` | `IntField` | IMU sample counts per unit |
| `n_gt_poses`, `n_odom_poses` | `IntField` | Ground-truth and wheel-odometry pose counts in the window |
| `n_metadata_messages` | `IntField` | Total messages across all `/metadata/*` numeric telemetry streams |
| `has_image`, `has_pointcloud`, `has_imu`, `has_gt_pose`, `has_audio` | `BooleanField` | Capability flags derived from the stream counts above, for filtering episodes without opening the `.mcap` |
| `changelog` | `StringField` | Verbatim contents of the source session's `CHANGELOG.md`, if present |

### MCAP topics (inside each episode)

Each `.mcap` file contains ~30 topics, all logged with Foxglove-native schemas
(chosen for compatibility with FiftyOne's multimodal viewer's Image/3D/Map/Plot/
Message tiles):

| Topic(s) | Schema | Tile | Notes |
|----------|--------|------|-------|
| `/basler/image`, `/zedx_left/image`, `/zedx_right/image` | `foxglove.CompressedImage` (png) | Image | ZED-X images have their alpha channel stripped before re-encoding (it carried no signal — verified near-constant ≈255 with no correlation to luminance) |
| `/basler/calibration`, `/zedx_left/calibration`, `/zedx_right/calibration` | `foxglove.CameraCalibration` | (enables reprojection) | Static, logged once from the source `calib/*.json` |
| `/robosense/points`, `/leishen/points` | `foxglove.PointCloud` | 3D | `x,y,z,intensity` (float32) + `ring` (uint16); the per-point absolute timestamp field present in the raw `.bin` files is dropped (the per-scan `log_time` already carries scan time) |
| `/navtech/image_polar` | `foxglove.RawImage` (mono8) | Image | Raw polar radar scan, header columns (per-azimuth timestamp/encoder) stripped |
| `/navtech/image_bev` | `foxglove.RawImage` (mono8) | Image | Cartesian bird's-eye-view radar conversion, using the source devkit's own official `polar_to_cartesian()` formula |
| `/audio_left/audio`, `/audio_right/audio` | `foxglove.RawAudio` (pcm-s16) | Message | 44.1 kHz mono 1-second clips |
| `/vectornav/imu`, `/xsens/imu` | generic JSON | Plot | `{wx,wy,wz,ax,ay,az}` per sample |
| `/gt_pose` | `foxglove.PoseInFrame` | 3D | PPK-GNSS ground-truth pose, `frame_id="map"`, recentered by subtracting the episode's first pose (see `gt_origin_utm`); orientation is always the identity quaternion (the source PPK pipeline provides no reliable attitude estimate) |
| `/tf` | `foxglove.FrameTransforms` (dynamic) | (enables 3D) | `map`→`base_link`, one transform per ground-truth pose — drives correct 3D placement of the moving robot over time |
| `/tf_static` | `foxglove.FrameTransforms` (static) | (enables 3D) | Full sensor-extrinsic tree rooted at `base_link` (`base_link`→`{footprint, robosense}`, `robosense`→`{basler, zedx_left, navtech, leishen}`, `zedx_left`→`{zedx_right, vectornav, xsens}`) |
| `/odom` | `foxglove.Odometry` | 3D + Plot | Wheel odometry (ideal differential-drive model), `frame_id="odom"` |
| `/metadata/*` (16 streams: battery, current/voltage/velocity ×2, cmd-velocity ×3, camera brightness/exposure/PTP status, barometer, IMU pressure/temperature, weather, snow depth) | generic JSON | Plot | Numeric telemetry, windowed to the episode's 30 seconds |

### Label types and why

There are **no object/segmentation/classification labels** — the source FoMo
dataset ships none (verified directly, not assumed from the paper's abstract).
The closest thing to a "label" is the **ground-truth pose trajectory**
(`/gt_pose` + dynamic `/tf`), mapped to `foxglove.PoseInFrame` rather than a
FiftyOne `Detection`/`Classification` type because it is inherently a
continuous, time-varying multimodal stream, not a per-sample annotation — this
is also why it lives inside the `.mcap` rather than as a separate FiftyOne label
field. Note that including ground truth in the `.mcap` at all is an addition
beyond the source devkit's own official converter, which omits it entirely
(their pipeline treats `gt.txt` as an offline-evaluation-only artifact).

### `dataset.info`

The dataset-level `info` dict records the original S3 source, paper citation,
license, and a short note explaining the sampling/authoring approach (30-second
trimmed episodes, 30 GB budget, one per trajectory type across 6 deployments).

### Parsing decisions

- **Lidar `.bin` point struct is 26 bytes/point, not the paper/tutorial's stated
  6×float32 (24 bytes)**: `x,y,z,intensity` (float32) + `ring` (uint16) +
  `timestamp` (uint64, µs) — discovered empirically by byte-stride search, then
  independently confirmed against the source devkit's own Python loader.
- **ZED-X PNGs are RGBA, not RGB** — alpha is near-constant noise, dropped
  before logging.
- **GT/odom coordinates**: `gt.txt` positions are absolute UTM-like coordinates;
  recentered per episode by subtracting the first pose (stored in
  `gt_origin_utm` for reversibility). Wheel odometry is already
  episode-relative, so it is logged as-is.
- **`odom.csv` has a duplicated, broken header** (`tax,tay,taz` appears twice;
  the second trio is always empty) — only the first 14 real columns are parsed.
- **Radar** (`navtech/*.png`): decoded per the Oxford radar-dataset convention
  (8-byte per-azimuth timestamp + 2-byte encoder + 1 reserved byte + range
  bins), verified against the source devkit; both the raw polar form and a
  Cartesian BEV conversion (using the devkit's exact official formula) are
  included.
- **Metadata CSVs** (`meteo_data.csv`, `snow_data.csv`) have a 3-row header
  (name/unit/stat-type) — the two extra rows are skipped when parsing.
- **Episode window anchoring**: each 30-second window starts at the first real
  lidar frame timestamp, not the GT log start — in one deployment the GT logger
  started up to 56 seconds before the sensors, which would otherwise produce an
  empty window.
- **Episode/session selection**: from the 60 available sessions, 6 were chosen
  to maximize both trajectory-type coverage (all 6: red, blue, green, magenta,
  yellow, orange) and seasonal/deployment diversity (6 distinct dates spanning
  Nov 2024–Sep 2025), while keeping the total raw+authored footprint under a
  30 GB budget.

## Dataset Creation

### Curation Rationale

The source FoMo dataset (9.43 TB, 60 sessions) is far too large to explore
directly, and its raw per-file format (individual PNG/`.bin`/`.csv` files per
sensor per timestamp) is not natively viewable in FiftyOne. This sample exists to
(a) demonstrate that FoMo's raw sensor format can be authored into FiftyOne's
`.mcap`-based multimodal viewer, and (b) provide a small, diverse, quickly
downloadable sample — one episode per trajectory type across 6 distinct seasonal
deployments — for exploring the dataset's structure and multi-season character
without downloading the full source dataset.

### Source Data

#### Data Collection and Processing

The original FoMo data was collected using a Clearpath Warthog uncrewed ground
vehicle equipped with two lidars (RoboSense Ruby Plus, Leishen LS128S1), an FMCW
radar (Navtech CIR-304H), a stereo camera (ZED X) and a monocular camera (Basler
ace2), two IMUs (VectorNav VN100, Xsens MTi-30), two microphones, and three rover
plus one static GNSS receiver for PPK ground-truth generation. Recording took
place across 12 deployments between November 2024 and October 2025 in Forêt
Montmorency, spanning conditions from -19°C with over 1 m of snow to 18°C summer
days. Ground truth was generated by fusing the three rover GNSS trajectories via
a point-to-Gaussian optimization against a CORS-corrected static base station
(see the source paper's "Ground Truth" section for the full pipeline).

For this FiftyOne sample specifically: 6 of the 60 sessions were selected (see
Parsing decisions above), each trimmed to its first 30 seconds of synchronized
sensor data, then authored into a single `.mcap` file per episode using
`foxglove-sdk`. Every raw stream was parsed directly from source bytes (not
assumed from the paper or tutorial documentation) and cross-checked against the
source devkit's own Python/Rust loaders where available.

#### Who are the source data producers?

The original sensor data was collected by researchers at Norlab (Université
Laval) and the University of Toronto Robotics Institute, operating the UGV
during field deployments in Forêt Montmorency, Quebec, Canada.

### Annotations

#### Annotation process

None. The source FoMo dataset contains no manual or model-generated
object/segmentation/classification annotations of any kind. The only
ground-truth signal is the PPK-GNSS-derived robot pose trajectory
(`gt.txt`/`gt_covariance.csv`), produced via automated post-processing of the
recorded GNSS receiver data (RINEX + PPK in Emlid Studio, followed by a
point-to-Gaussian multi-receiver fusion), included in this sample as the
`/gt_pose` and `/tf` (map→base_link) streams in each episode's `.mcap`.

#### Who are the annotators?

Not applicable — no manual annotation was performed. The ground-truth pose
trajectory was generated automatically by the original FoMo authors' PPK-GNSS
post-processing pipeline.

#### Personal and Sensitive Information

Per the source paper, camera images were manually anonymized for faces and
license plates using a custom-built pipeline before public release. No other
personal or sensitive information is present — the recordings are sensor
telemetry from an unpopulated boreal forest and gravel/paved access roads.

## Citation

**BibTeX:**

```bibtex
@misc{Boxan2026_fomo,
  title   = {{FoMo: A Multi-Season Dataset for Robot Navigation in For\^et Montmorency}},
  author  = {Matěj Boxan and Gabriel Jeanson and Alexander Krawciw and Effie Daum
             and Xinyuan Qiao and Sven Lilge and Timothy D. Barfoot and François Pomerleau},
  year    = {2026},
  eprint  = {2603.08433},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url     = {https://arxiv.org/abs/2603.08433}
}
```

**APA:**

Boxan, M., Jeanson, G., Krawciw, A., Daum, E., Qiao, X., Lilge, S., Barfoot, T.
D., & Pomerleau, F. (2026). *FoMo: A Multi-Season Dataset for Robot Navigation in
Forêt Montmorency*. arXiv:2603.08433.

## More Information

This is a 6-of-60-session, 30-seconds-per-session sample of the full FoMo
dataset, selected under a 30 GB size budget — it is not a substitute for the full
9.43 TB source dataset for benchmarking purposes. The full dataset (12
deployments, 60 sessions, six trajectory types, one year of seasonal coverage)
is publicly available, unsigned, at `s3://fomo-dataset`.

## Dataset Card Authors

[Harpreet Sahota](https://huggingface.co/harpreetsahota)