README.md

---
annotations_creators:
- machine-generated
- expert-generated
language:
- en
license: cc-by-4.0
size_categories:
- 10K<n<100K
task_categories:
- object-detection
- image-classification
pretty_name: AppleGrowthVision
tags:
- fiftyone
- group
- object-detection
- image-classification
- agriculture
- precision-agriculture
- phenology
- bbch
- stereo
- apples
dataset_summary: '

This is a [FiftyOne](https://github.com/voxel51/fiftyone) dataset with 11,397
groups (21,407 samples) of apple orchard imagery, curated with the [FiftyOne
Dataset Curation skill](https://github.com/voxel51/fiftyone).

## 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/AppleGrowthVision")

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

'
---

# Dataset Card for AppleGrowthVision

![image/png](applegrowth.gif)

This is a [FiftyOne](https://github.com/voxel51/fiftyone) dataset with 11,397
groups (21,407 samples) of apple orchard imagery.

## 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/AppleGrowthVision")

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

## Dataset Details

### Dataset Description

AppleGrowthVision is a large-scale, longitudinal dataset of apple orchard
imagery collected at two German field sites across one or more full growth
cycles. It combines two capture modalities: a fixed, fully-calibrated stereo
rig (Brandenburg / "BB Obst" farm, Wesendahl) and handheld smartphone imagery
(Saxony / Pillnitz orchard, near Dresden). Unlike prior apple-detection
datasets, every image is tied to a specific point on the extended BBCH scale
for pome fruit (bud development through senescence), and a subset of images
is densely annotated with apple bounding boxes. Ten Pillnitz trees also carry
a manually verified (non-destructive) reference apple count, giving a small
yield-estimation ground truth in addition to the detection labels.

- **Curated by:** Laura-Sophia von Hirschhausen, Jannes S. Magnusson, Mykyta
  Kovalenko, Fredrik Boye, Tanay Rawat, Peter Eisert, Anna Hilsmann, Sebastian
  Pretzsch, Sebastian Bosse (Fraunhofer HHI and Fraunhofer IVI)
- **Funded by:** German Federal Ministry for Economic Affairs and Climate
  Action, NaLamKI project (Grant 01MK21003D); German Federal Ministry of Food
  and Agriculture (BMEL), LANDNETZ project (Grant 28DE101C18)
- **Shared by:** Fraunhofer HHI / Fraunhofer IVI (original data); this
  FiftyOne conversion shared by Harpreet Sahota
- **Language(s):** N/A (imagery dataset with English-language metadata)
- **License:** CC BY 4.0

### Dataset Sources

- **Repository:** https://fraunhofer-hhi.github.io/AppleGrowthVision /
  https://datacloud.hhi.fraunhofer.de/s/KLFXDw9cWSzXk95
- **Paper:** [AppleGrowthVision: A large-scale stereo dataset for
  phenological analysis, fruit detection, and 3D reconstruction in apple
  orchards](https://arxiv.org/abs/2505.14029) (arXiv:2505.14029)
- **Demo:** [More Information Needed]

## Uses

### Direct Use

- Apple fruit detection / object detection benchmarking (single class,
  `apple`), including combining this data with MinneApple and the Monastery
  Apple Dataset (MAD) as done in the source paper.
- BBCH principal growth-stage classification (bud development, leaf
  development, shoot development, inflorescence emergence, flowering, fruit
  development, fruit/seed maturity, senescence).
- Calibrated multi-view stereo reconstruction of orchard scenes (Brandenburg
  subset only, using the per-date camera calibration stored in
  `dataset.info` / the companion camera-rig visualization dataset).
- Yield-estimation research, using the `num_apples` detection counts and the
  10-tree manual `reference_apple_count` ground truth as a small validation
  set.

### Out-of-Scope Use

- The dataset is not exhaustively annotated: only 1,457 of the 21,407 images
  (~6.8%) have `ground_truth` apple bounding boxes, concentrated in
  BBCH principal stages 5-9 (fruit-visible stages) and specific capture
  dates/rows. It should not be treated as a fully-labeled detection set --
  filter to the `annotated` or `has-apples` saved views (see below) for
  detection work.
- Not intended for tasks involving human subjects; this is strictly
  agricultural/orchard imagery.

## Dataset Structure

This is a **grouped** FiftyOne dataset (`media_type="image"` in every slice)
with **3 group slices**, all populated from the same underlying sample
collection:

| Slice | Samples | Site | Description |
|-------|---------|------|-------------|
| `left` (default) | 10,010 | Brandenburg | Left camera of the calibrated stereo rig |
| `right` | 10,010 | Brandenburg | Right camera of the calibrated stereo rig |
| `image` | 1,387 | Saxony/Pillnitz | Handheld smartphone shots, one sample per group (singleton groups, reusing the group mechanism so both sites live in one dataset) |

Total: 11,397 groups, 21,407 flat samples.

### Fields

| Field | FiftyOne type | Description |
|-------|---------------|-------------|
| `filepath` | `StringField` | Path to the image (verbatim from source) |
| `tags` | `ListField(StringField)` | Provenance tags, e.g. `brandenburg`, `saxony`, `2023`, `2024`, `pillnitz_2024_backup`, `row36`-`row47`, `labeled`, `has_apples` |
| `metadata` | `Metadata` | Standard FiftyOne image metadata (size, mime type, dimensions) |
| `group` | `Group` | Group field (`group.name` is the slice: `left`/`right`/`image`) |
| `site` | `StringField` | `brandenburg` or `saxony` |
| `subset` | `StringField` | Finer-grained data-drop identifier: `brandenburg`, `saxony_2023`, `saxony_2024`, `saxony_2024_counted_trees`, `saxony_pillnitz_row{36,37,38,39,40,44,45,46,47}` |
| `capture_date` | `StringField` | Capture date (`YYYY-MM-DD`), set on Brandenburg samples from the per-date data folder |
| `capture_datetime` | `StringField` | Capture date + time (`YYYYMMDD_HHMMSS`), set on Saxony samples from the filename timestamp |
| `shot_index` | `IntField` | Brandenburg shot index within a capture date (shared by the L/R pair) |
| `camera_hangle` / `camera_kappa0` / `camera_aspect` | `FloatField` | Per-camera calibration parameters (Brandenburg only), parsed from the date's calibration XML |
| `calib_variant` | `StringField` | `"after"` when a date had both a before/after calibration file (the "after" file is used as canonical) |
| `row_index` / `tree_index` | `IntField` | Orchard row and tree-within-row index (Saxony only) |
| `picture_index` | `IntField` | Picture number for a given tree/date (Saxony 2023/2024 main drops only) |
| `is_detail` | `BooleanField` | Whether the filename was flagged as a `_detail` (close-up) shot (Saxony 2023 only) |
| `growth_stage` | `StringField` | Free-text growth stage from the 2023 filenames: `blossom`, `small_fruit`, `middle_fruit`, `fruit` |
| `bbch_code` | `StringField` | Extended BBCH growth-stage code (e.g. `"65"`, or `"71/72"` for one ambiguous table entry), derived from a site+date lookup table transcribed from the paper's Table 2 |
| `bbch_principal_stage` | `IntField` | First digit of `bbch_code` (the BBCH *principal* stage, 0-9) |
| `bbch_stage_name` | `StringField` | Human-readable principal-stage name (e.g. `"Flowering"`), from the paper's Table 1 |
| `ground_truth` | `Detections` | Apple bounding boxes (single class `apple`), converted from the source COCO annotations. Only present on samples that were part of the paper's dense-annotation pass; may be an *empty* `Detections` for reviewed-but-empty images (see below) |
| `num_apples` | `IntField` | `len(ground_truth.detections)`, set alongside every `ground_truth` (derived convenience field for counting/sorting) |
| `reference_apple_count` | `IntField` | Manually verified (non-harvested) apple count for 10 specific Pillnitz trees, from the paper's reference-count table |
| `reference_apple_count_bad_quality` | `IntField` | Additional apples noted as "bad quality" in that same manual count, set alongside `reference_apple_count` (0 when not noted) |

### Label type

The only label type is `Detections` (`ground_truth`), a single class,
`apple`, with relative `[x, y, w, h]` bounding boxes, plus per-box `area` and
`iscrowd` carried over from the source COCO annotations. Bounding boxes (not
polygons/masks) were used because that is the annotation format released
with the source data.

`labeled` (in `tags`) marks every sample that appears in a source COCO file,
including reviewed images with zero boxes (e.g. blossom-stage close-ups with
no visible fruit). `has_apples` marks the subset of those with at least one
box. This distinguishes "reviewed, no apples visible" from "never reviewed"
(for the latter, `ground_truth` is simply unset).

### `dataset.info`

- `description`, `source` (Fraunhofer HHI data-cloud share), `paper`
  (arXiv id)
- `paper_reported_stats`: the paper's headline numbers (9,317 stereo images,
  1,125 densely annotated images, 31,084 apple labels) alongside a note that
  the actual imported counts differ somewhat (10,010 Brandenburg stereo
  pairs on disk; 1,457 densely-annotated images across both sites) --
  documented rather than silently reconciled, since the discrepancy is in
  the released data itself, not in the FiftyOne conversion.
- `bbch_scale`: a short legend for the `bbch_code` / `bbch_principal_stage`
  taxonomy (Table 1 of the paper).

### Saved views

The dataset ships with 7 saved views for common access patterns:
`brandenburg-stereo-pairs`, `saxony-images`, `annotated`, `has-apples`,
`reference-count-trees`, `fruit-visible-stages` (BBCH principal stages 5-9),
and `phenology-timeline` (sorted by BBCH principal stage, then capture date).

### Indexes

Indexes exist on `site`, `subset`, `capture_date`, `growth_stage`,
`bbch_code`, `bbch_principal_stage`, `bbch_stage_name`, `row_index`,
`tree_index`, `num_apples`, `reference_apple_count`, `tags`, plus compound
indexes on `(row_index, tree_index)` and `(site, bbch_principal_stage)`.

### Not included in this repository

A companion 3D FiftyOne dataset (`AppleGrowthVision-camera-rig`) visualizes
the Brandenburg stereo rig geometry as 16 `.fo3d` scenes (one per calibrated
capture date), reconstructed from the calibration XML extrinsics. It is kept
as a separate, small local dataset rather than a group slice here, since the
rig geometry is per-date (16 configs), not per-image, and is not part of
this Hub repository.

## Dataset Creation

### Curation Rationale

Existing apple-orchard datasets (MinneApple, the Washington State University
robotic-harvesting dataset, the Michigan State O2RNet dataset, and the
Monastery Apple Dataset) advanced fruit detection but did not capture growth
stage or stereo/3D structure, and typically covered a limited window within
a single season. AppleGrowthVision was curated to capture a complete
phenological growth cycle with expert-validated BBCH stage labels and
calibrated stereo imagery, enabling phenological analysis, fruit detection,
and 3D orchard reconstruction from the same data.

### Source Data

#### Data Collection and Processing

Two field sites were imaged:

- **Brandenburg ("BB Obst"), Wesendahl, near Berlin:** the same 33 Jonagold
  trees were imaged on 18 occasions throughout 2022-2023 using two
  calibrated Canon EOS 550D cameras (20mm lenses) triggered simultaneously,
  from front, left, right, and steep-angle views plus a circular sweep
  around the first tree in the row.
- **Pillnitz orchard (LfULG Sachsen), near Dresden:** rows ~70m long, trees
  spaced 1m apart in spindle form, 3m between rows. A subset of trees was
  photographed from multiple angles with a smartphone camera across 2023
  (blossom through ripe fruit) and revisited in 2024 (single growth stage,
  after the orchard suffered fruit loss to late frost that year).
  Additional rows were covered by walking each row in one direction and back
  along the other side, capturing front/back views per tree.

Each Brandenburg capture date has its own camera calibration (extrinsics +
intrinsics), stored as XML and used here to populate the per-sample
`camera_hangle`/`camera_kappa0`/`camera_aspect` fields (and to build the
separate camera-rig 3D visualization dataset).

#### Who are the source data producers?

Imagery was collected by the paper's authors and collaborators at
Fraunhofer HHI and Fraunhofer IVI, at the BB Obst commercial orchard
(Wesendahl) and the LfULG Sachsen research/education orchard (Pillnitz).

### Annotations

#### Annotation process

Apple bounding-box annotation used an in-house annotation tool combining
AI-assisted pre-labeling with human verification: a YOLOv8 model was first
trained on a small manually-labeled seed set (108 non-stereo + 70 stereo
images, combined with the full MinneApple dataset) and used to auto-annotate
the remaining images. Annotations were only generated for images in BBCH
principal stages 5-9 (inflorescence emergence through senescence), since
earlier stages have no fruit to detect. Human annotators then reviewed and
corrected the AI-generated boxes for 70 stereo images and 777 non-stereo
images; final boxes are stored in Darknet/YOLO format and converted to COCO
for release (and to FiftyOne `Detections` here).

Growth-stage labeling (the BBCH principal stage per capture date) was
performed manually by an expert from LfULG Sachsen on a randomly-selected
subset of images per date. The reference apple counts for 10 Pillnitz trees
were obtained by manual visual counting in the field, without harvesting.

#### Who are the annotators?

Apple bounding boxes: the paper's authors/team, using the semi-automated
AI-assisted pipeline described above. BBCH growth-stage labels: an
agricultural expert from LfULG Sachsen (Sächsisches Landesamt für Umwelt,
Landwirtschaft und Geologie). Reference apple counts: field staff at the
Pillnitz orchard.

#### Personal and Sensitive Information

None. The dataset contains only orchard/tree/fruit imagery; no people or
personal data are depicted or recorded.

## Citation

**BibTeX:**

```bibtex
@article{vonhirschhausen2025applegrowthvision,
  title   = {AppleGrowthVision: A large-scale stereo dataset for phenological
             analysis, fruit detection, and 3D reconstruction in apple
             orchards},
  author  = {von Hirschhausen, Laura-Sophia and Magnusson, Jannes S. and
             Kovalenko, Mykyta and Boye, Fredrik and Rawat, Tanay and
             Eisert, Peter and Hilsmann, Anna and Pretzsch, Sebastian and
             Bosse, Sebastian},
  journal = {arXiv preprint arXiv:2505.14029},
  year    = {2025}
}
```

**APA:**

von Hirschhausen, L.-S., Magnusson, J. S., Kovalenko, M., Boye, F., Rawat,
T., Eisert, P., Hilsmann, A., Pretzsch, S., & Bosse, S. (2025).
AppleGrowthVision: A large-scale stereo dataset for phenological analysis,
fruit detection, and 3D reconstruction in apple orchards. *arXiv preprint
arXiv:2505.14029*.

## More Information

Extending MinneApple with AppleGrowthVision improved YOLOv8 F1-score by
7.69%, and adding it to MinneApple + MAD improved Faster R-CNN F1-score by
31.06% (paper Section 4.1). Six BBCH principal stages were predicted with
over 95% accuracy using VGG16, ResNet152, DenseNet201, and MobileNetv2
(paper Section 4.2). The Brandenburg calibrated stereo imagery was also used
for multi-view stereo orchard reconstruction via DISK features + LightGlue +
COLMAP (paper Section 4.3); this repository does not include the
reconstructed point clouds/meshes themselves, only the source imagery and
calibration.

## Dataset Card Authors

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