WildBe-v2: Expanded Drone Imagery of Wild Berries in Finnish Forests

Example Images

Introduction

We present WildBe-v2, the evolution and expansion of the original WildBe dataset, the first drone-captured collection of wild berries from Finnish peatlands and forest canopies. WildBe-v2 introduces 11 classes, including five berry types (bilberries, cloudberries, crowberries, lingonberries, and bog bilberries) annotated in ripe and unripe states, plus the mushroom class. The dataset now comprises 21K images and over 61K annotations collected from seven distinct locations. This broader coverage enhances variability in lighting, vegetation, and background complexity, offering a richer benchmark for research on object detection and visual understanding in the wild.

Dataset Classes

The dataset contains 11 classes, mapped to the integer IDs in the labels:

  0  →  bilberry 
  1  →  bilberry-unripe 
  2  →  cloudberry 
  3  →  cloudberry-unripe  
  4  →  crowberry  
  5  →  crowberry-unripe  
  6  →  lingonberry  
  7  →  lingonberry-unripe  
  8  →  bog-bilberry 
  9  →  bog-bilberry-unripe
 10  →  mushroom

How to use: an example of visualization

import json

from datasets import load_dataset
from PIL import ImageDraw

# Color map for classes
classes_color_map = {
    0: (4, 200, 86),     # dark-pastel-green
    1: (14, 251, 113),   # spring-green
    2: (204, 153, 0),    # goldenrod
    3: (255, 199, 31),   # mikado-yellow
    4: (0, 143, 204),    # blue-ncs
    5: (31, 188, 255),   # deep-sky-blue
    6: (184, 0, 95),     # amaranth-purple
    7: (255, 10, 137),   # magenta
    8: (255, 181, 112),  # sandy-brown
    9: (255, 202, 153),  # peach
    10: (165, 82, 42),   # rustbrown
}

# Load the dataset
dataset = load_dataset("FBK-TeV/WildBe-v2", split="validation")
dataset = dataset.shuffle(seed=42)

sample_idx = 30
image = dataset[sample_idx]["image"]
labels = json.loads(dataset[sample_idx]["labels"])

draw = ImageDraw.Draw(image)

for label in labels:
    center_x = label["x"] * dataset[sample_idx]["w"]
    center_y = label["y"] * dataset[sample_idx]["h"]
    width = label["width"] * dataset[sample_idx]["w"]
    height = label["height"] * dataset[sample_idx]["h"]
    draw.rectangle(
        [
            (center_x - width / 2, center_y - height / 2),
            (center_x + width / 2, center_y + height / 2),
        ],
        outline=classes_color_map[label["class"]],
        width=2,
    )

image.show()

Data Fields

index: An integer representing the unique identifier for each example.
image: A PIL image.
w: image width
h: image height
split: A string indicating the data split, e.g., 'train', 'validation', or 'test'.
provenance: A string ("A", "B", ...) indicating the area in which the image has been collected
source_image_id: A string representing the unique identifier for the source image from which the image was cropped. 
labels: A list of dictionaries, each containing:
  class: An integer representing the class identifier.
  label: A string representing the class name.
  x: A float representing the normalized x-coordinate of the center of the bounding box.
  y: A float representing the normalized y-coordinate of the center of the bounding box.
  width: A float representing the normalized width of the bounding box.
  height: A float representing the normalized height of the bounding box.

Acknowledgement

The FEROX project has received funding from the European Union’s Horizon Framework Programme for Research and Innovation under the Grant Agreement no 101070440 - call HORIZON-CL4-2021-DIGITAL-EMERGING-01-10: AI, Data and Robotics at work (IA).

Partners