---
license: cc-by-4.0
task_categories:
  - image-segmentation
tags:
  - fiftyone
  - image
  - segmentation
  - agriculture
  - rgb-d
  - depth
  - crop-weed
  - weed-detection
  - precision-agriculture
size_categories:
  - 1K<n<10K
---

# Dataset Card for WE3DS

This is a [FiftyOne](https://github.com/voxel51/fiftyone) dataset with 2568 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
dataset = load_from_hub("Voxel51/WE3DS")

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

## Dataset Details

![image/png](we3ds.gif)

### Dataset Description

WE3DS is an RGB-D image dataset for multi-class plant species semantic segmentation in
crop farming. It contains 2,568 densely, hand-annotated RGB-D images (color image +
stereo-derived distance map) captured under natural outdoor light conditions at the
experimental farm of the University of Natural Resources and Life Sciences, Vienna
(BOKU) in Groß-Enzersdorf, Austria, across 25 measurement dates in 2020 and 2021.

Each image is paired with a dense, per-pixel ground-truth segmentation mask covering
soil, 7 crop species, and 10 weed species (18 "real" classes plus a `void` class for
uncertain/unlabeled regions), along with a raw and a hole-filled ("refined") distance
map, and per-image acquisition metadata (GPS location, timestamp, weather, wind,
seeding date, and measured plant/canopy height).

- **Curated by:** Florian Kitzler, Norbert Barta, Reinhard W. Neugschwandtner, Andreas Gronauer, Viktoria Motsch (University of Natural Resources and Life Sciences, Vienna — BOKU)
- **Funded by:** The "DiLaAg — Digitalization and Innovation Laboratory in Agricultural Sciences" project, supported by the Government of Lower Austria and the private foundation Forum Morgen
- **Shared by:** Harpreet Sahota (FiftyOne-formatted version)
- **Language(s):** N/A (image dataset)
- **License:** CC BY 4.0

### Dataset Sources

- **Repository (original data + code):** https://doi.org/10.5281/zenodo.7457983
- **Paper:** Kitzler, F.; Barta, N.; Neugschwandtner, R.W.; Gronauer, A.; Motsch, V. WE3DS: An RGB-D Image Dataset for Semantic Segmentation in Agriculture. *Sensors* 2023, 23, 2713. https://doi.org/10.3390/s23052713

## Uses

### Direct Use

- Training and benchmarking RGB, depth-only, or RGB-D semantic segmentation models for
  crop/weed discrimination in early growth stages.
- Studying the effect of fusing distance (depth) information with RGB imagery for
  outdoor agricultural scene understanding.
- Analyzing plant/weed distribution, canopy height, or acquisition-condition (weather,
  wind, date) effects on segmentation performance.

### Out-of-Scope Use

- The dataset covers a single experimental farm site (Groß-Enzersdorf, Austria) and a
  fixed set of seeded crop/weed species in early growth stages only — it is not
  representative of other geographies, growth stages, or cropping systems, and models
  trained on it should not be assumed to generalize without further validation.

## Dataset Structure

This is a flat `image`-type FiftyOne dataset (no grouping) with 2,568 samples and a
`train`/`test` split carried over from the original release (1,540 train / 1,028 test,
a ~60/40 split), stored as sample tags.

| Field | FiftyOne type | Description |
|---|---|---|
| `filepath` | `StringField` | Path to the RGB image (`images/img_XXXXX.png`, 1600×1144, uint8) |
| `tags` | `ListField` | `train` or `test`, from the dataset's official split files |
| `ground_truth` | `Segmentation` | Dense per-pixel class-index mask (19 mask values: `void` + soil + 7 crops + 10 weeds); `mask_targets` stored on the dataset (see below) |
| `depth_raw` | `Heatmap` | Raw stereo-derived distance map (16-bit, same filename as the RGB image). `0` marks pixels where stereo matching failed due to occlusion or insufficient surface texture. Values are in units of 0.1 mm (e.g. a pixel value of 7268 ≈ 726.8 mm). `range` is set per-sample to `[0, 99th percentile of nonzero pixels]` to keep a handful of stereo-matching sentinel/noise pixels from washing out the visualized contrast |
| `depth_refined` | `Heatmap` | Hole-filled/interpolated version of `depth_raw` with no zero/invalid pixels. `range` is set per-sample to `[min, max]` of that image for full local contrast |
| `location` | `GeoLocation` | Capture GPS point (`[lon, lat]`). Unset (`None`) for 40 samples where the source GNSS reading was missing (see note below) |
| `lat`, `lon` | `FloatField` | Capture latitude/longitude (also duplicated into `location` above). Unset for the same 40 samples |
| `date`, `time`, `utc` | `StringField` | Local capture date/time and UTC timestamp, as recorded by the acquisition system |
| `weather` | `StringField` | Subjective weather condition at capture time: `sunny`, `cloudy`, or `mixed` |
| `wind` | `StringField` | Subjective wind intensity at capture time: `light`, `medium`, or `strong` |
| `seeding_date` | `StringField` | Date the parcel was seeded |
| `height_mm` | `IntField` | Measured plant/canopy height (mm) at time of capture |

**Label type notes:**

- Segmentation masks are stored as `fo.Segmentation` rather than per-instance
  `Detections`/polygons, matching the source data (single-channel class-index PNGs).
  The original annotation process (CVAT, polygon-per-plant-instance) produced instance
  data, but the distributed masks are the rasterized, class-indexed result.
- Depth maps are stored as `fo.Heatmap` (dense per-pixel float/int maps) rather than as
  plain auxiliary image fields, so they render as color overlays in the FiftyOne App.
  Because the true depth range varies substantially per image (capture height was
  adjusted over the course of the field trials), a single fixed range across the whole
  dataset made the heatmaps look nearly uniform — ranges are therefore computed
  **per sample** as described in the field table above.
- The source `info.csv` has the literal string `"nan"` for `lat`/`lon`/`utc` on 40
  samples (three short GNSS dropout windows during acquisition on `10.06.2021`,
  `img_01179`–`img_01208` and `img_01229`–`img_01238`). Rather than storing an invalid
  `GeoLocation` point (which breaks map-based visualization), these fields are left
  unset (`None`) on the affected samples.

**`dataset.info` contents:**

```python
{
    "description": "WE3DS: an RGB-D image dataset for semantic segmentation of crops and weeds in agricultural fields (Kitzler et al., Sensors 2023).",
    "source": "https://doi.org/10.3390/s23052713",
    "class_names": ["void", "soil", "broad bean", "corn spurry", "red-root amaranth",
                    "common buckwheat", "pea", "red fingergrass", "common wild oat",
                    "cornflower", "corn cockle", "corn", "milk thistle", "rye brome",
                    "soybean", "sunflower", "narrow-leaved plantain",
                    "small-flower geranium", "sugar beet"],
}
```

`dataset.mask_targets["ground_truth"]` maps the 19 mask pixel values (0–18) to these
same class names, in the same order (index 0 = `void`, 1 = `soil`, 2–18 = the 7 crop
and 10 weed species).

## Dataset Creation

### Curation Rationale

Publicly available RGB image datasets for agricultural semantic segmentation were
scarce and typically lacked additional distance (depth) information, despite evidence
from other domains (indoor scenes, street scenes) that fusing RGB with depth improves
segmentation quality. WE3DS was created to fill this gap: a real-world, natural-light,
multi-class (not just soil/crop/weed) RGB-D dataset for crop farming, together with a
benchmark showing that adding distance information does improve segmentation quality
in this domain as well.

### Source Data

#### Data Collection and Processing

Images were captured with a custom RGB-D sensor: two industrial XIMEA MC023CG-SY
cameras (2.3MP Sony IMX174LLJ-C sensor, global shutter) in a stereo setup with a 4-5cm
baseline and 12mm fixed-focal-length lenses (TAMRON M112FM12), achieving ~0.4mm/pixel
ground resolution and ~1.6mm depth accuracy at a 90cm working height. The sensor was
mounted top-down on a two-wheeled measurement trolley (60cm wheel spacing) alongside a
GNSS module (Emlid Reach M2), and images were captured at 1 FPS while the trolley was
pushed along crop rows.

Field trials ran over seven repetitions (four in 2020, three in 2021) across small
parcels (2.5m × 9m in 2020, 1.5m × 5m in 2021) at the BOKU experimental farm in
Groß-Enzersdorf, Austria, seeding 39 plant species across 247 parcels (a subset of which
were usable after excluding parcels with excessive weeds, poor emergence, or bird
damage). In total, 6,224 stereo image pairs were collected across 25 measurement dates
and 84 parcels; 2,568 were selected (excluding poor-quality or plant-absent images) for
ground-truth annotation and form this dataset. An additional 3,656 captured pairs were
left unannotated at publication time.

For each selected pair, the raw stereo images were rectified using per-measurement-date
camera calibration (HALCON `map_image`), then a distance map was computed via
normalized cross-correlation stereo matching (`binocular_distance`, 21-pixel template).
Non-overlapping border regions were cropped, producing the final 1600×1144 RGB image
and aligned distance map used here. Stereo matching fails for occluded or texture-poor
pixels (0.72% of pixels missing on average per image, up to 14% in the worst case) —
these are the `0`-valued pixels in `depth_raw`; `depth_refined` is the same map with
those holes filled by interpolation.

#### Who are the source data producers?

Images and metadata were collected by the paper's authors and collaborators at BOKU's
Institute of Agricultural Engineering and Institute of Agronomy, using their in-house
acquisition software and hardware described above.

### Annotations

#### Annotation process

Ground-truth annotation was performed in CVAT (Computer Vision Annotation Tool).
Annotators drew labeled polygons around each individual plant instance (of a known,
seeded species) in the unprocessed left RGB image. An additional class was used for
uncertain cases and small/unknown weeds overlooked during manual weed control; together
with species having too little annotation data, these were folded into the `void`
class, which is excluded from the paper's benchmark training/evaluation. The resulting
polygon annotations were then rectified and cropped to match the RGB-D image pair,
producing the dense, per-pixel class-index masks distributed with this dataset. Across
the dataset, 4,038 crop and 7,506 weed instances were annotated (average 4.5
plants/image: 1.6 crop + 2.9 weed instances, or ~38,234 plant pixels/image on average).

#### Who are the annotators?

Manual image annotation was performed by Nicole Burscha, Sarah Sierra Marroquin, and
Melvin Lukas Polleichtner (per the paper's acknowledgments).

#### Personal and Sensitive Information

None. The dataset contains agricultural field imagery (soil, crops, weeds) and
associated GPS/weather/timing metadata for a single non-residential experimental farm
site; no personal or sensitive information is present.

## Citation

**BibTeX:**

```bibtex
@article{kitzler2023we3ds,
  title     = {WE3DS: An RGB-D Image Dataset for Semantic Segmentation in Agriculture},
  author    = {Kitzler, Florian and Barta, Norbert and Neugschwandtner, Reinhard W. and Gronauer, Andreas and Motsch, Viktoria},
  journal   = {Sensors},
  volume    = {23},
  number    = {5},
  pages     = {2713},
  year      = {2023},
  publisher = {MDPI},
  doi       = {10.3390/s23052713}
}
```

**APA:**

Kitzler, F., Barta, N., Neugschwandtner, R. W., Gronauer, A., & Motsch, V. (2023). WE3DS: An RGB-D Image Dataset for Semantic Segmentation in Agriculture. *Sensors*, 23(5), 2713. https://doi.org/10.3390/s23052713

## More Information

The original dataset and modified ESANet benchmark code are available at
https://doi.org/10.5281/zenodo.7457983. In the source paper's benchmark, an RGB-D
ESANet model trained at full input resolution (1280×960) reached 70.7% mIoU across 18
classes (soil + 17 plant species), compared to 70.1% for RGB-only and 48.5% for
distance-only at the same resolution — confirming that the additional distance channel
improves segmentation quality in this domain.

## Dataset Card Authors

Harpreet Sahota (FiftyOne conversion and dataset card)