Wild Berry image dataset collected in Finnish forests and peatlands using drones

TL;DR

The paper tackles automated detection of wild berries to support safer, more efficient harvesting in Finland using drone imagery. It introduces WildBe, a drone-captured dataset of bilberries, cloudberries, crowberries, and lingonberries in peatlands and forest canopies, comprising 3,516 images and 18,336 bounding boxes annotated in YOLO format across four classes, collected from multiple sensors. The authors evaluate six detectors (Faster R-CNN, VarifocalNet, GLIP, DINO, ObjectBox, YOLOv8) under single-class and multi-class settings and across transfer scenarios (across forest areas, cameras, and a cross-dataset CRAID test) using COCO AP metrics; GLIP consistently yields top performance, though substantial domain gaps appear in cross-dataset transfer. WildBe is publicly available on HuggingFace, offering a valuable resource to develop robust berry-detection and domain-adaptation methods for challenging forest environments, with noted limitations and future work on expert fine-grained annotation.

Abstract

Berry picking has long-standing traditions in Finland, yet it is challenging and can potentially be dangerous. The integration of drones equipped with advanced imaging techniques represents a transformative leap forward, optimising harvests and promising sustainable practices. We propose WildBe, the first image dataset of wild berries captured in peatlands and under the canopy of Finnish forests using drones. Unlike previous and related datasets, WildBe includes new varieties of berries, such as bilberries, cloudberries, lingonberries, and crowberries, captured under severe light variations and in cluttered environments. WildBe features 3,516 images, including a total of 18,468 annotated bounding boxes. We carry out a comprehensive analysis of WildBe using six popular object detectors, assessing their effectiveness in berry detection across different forest regions and camera types. WildBe is publicly available on HuggingFace at https://huggingface.co/datasets/FBK-TeV/WildBe.

Figures (7)

• Figure 1: The WildBe dataset comprises images of wild berries, including bilberries, cloudberries, crowberries, and lingonberries, captured in peatlands and the undercanopy of Finnish forests using drones. Images are manually annotated with bounding boxes.
• Figure 2: WildBe statistics: (left-hand side) proportion of annotations for each class within the dataset and (right-hand side) distribution of bounding boxes across the images.
• Figure 3: Summary of the results quantified in terms of AP. (left-hand side) Radar visualisation that includes the evaluation of detectors in coping with single- and multi-class objects, transfer learning (TL) capabilities across different forest areas, TL abilities across different camera sensors, and across different datasets, specifically on CRAID Akiva2020craid. (right-hand side) Radar visualisation that includes the evaluation of detectors for each class of berries.
• Figure 4: Qualitative results. Columns show a comparison against the best-performing methods (first three columns) and the ground-truth reference (last column). Rows show different examples. Key -- Red: bilberry, azure: cloudberry, blue: crowberry, purple: lingonberry.
• Figure 5: Qualitative results for the cross-area transfer learning experiment. Columns: a comparison against the best-performing methods (first three columns) and the ground-truth reference (last column). Rows: different geographical areas: the first row contains an image captured in Area 1, while the second row shows an image collected in Area 2. Key -- Red: bilberry, blue: crowberry, purple: lingonberry.