AppleGrowthVision: A large-scale stereo dataset for phenological analysis, fruit detection, and 3D reconstruction in apple orchards
Laura-Sophia von Hirschhausen, Fraunhofer HHI, Jannes S. Magnusson ‡, Mykyta Kovalenko ‡, Fredrik Boye ‡, Tanay Rawat, Peter Eisert, Anna Hilsmann, Sebastian Pretzsch, Sebastian Bosse · 2025
Read AppleGrowthVision: A large-scale stereo dataset for phenological analysis, fruit detection, and 3D reconstruction in apple orchards by Laura-Sophia von Hirschhausen, Fraunhofer HHI, Jannes S. Magnusson ‡, Mykyta Kovalenko ‡, Fredrik Boye ‡, Tanay Rawat, Peter Eisert, Anna Hilsmann, Sebastian Pretzsch, Sebastian Bosse (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-14029.txt.
Introduction
Laura-Sophia von Hirschhausen and eight co-authors publish a large stereo image dataset of apple orchards across a growing season, in a 2025 paper. Precision farming models starve for realistic training data: orchards are dense, messy, and change weekly, and flat images cannot support three-dimensional reasoning about yield.
One part holds over ninety-three hundred high-resolution stereo pairs from a Brandenburg farm, spanning six validated growth stages over a full cycle. The other holds eleven hundred densely annotated images from two farms, with more than thirty-one thousand labeled apples. Stereo pairs enable depth reconstruction of trees, while growth-stage labels support timing analysis. The dataset extends an earlier public orchard collection and targets fruit detection, yield estimation, and structural modeling for orchard robotics.
Catalog record
| CREATOR | Laura-Sophia von Hirschhausen, Fraunhofer HHI, Jannes S. Magnusson ‡, Mykyta Kovalenko ‡, Fredrik Boye ‡, Tanay Rawat, Peter Eisert, Anna Hilsmann, Sebastian Pretzsch, Sebastian Bosse |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 27 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-14029
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
AppleGrowthVision: A large-scale stereo dataset for phenological analysis, fruit detection, and 3D reconstruction in apple orchards Laura-Sophia von Hirschhausen, Fraunhofer HHI, Jannes S. Magnusson ‡, Mykyta Kovalenko ‡, Fredrik Boye ‡, Tanay Rawat, Peter Eisert, Anna Hilsmann, Sebastian Pretzsch, Sebastian Bosse Abstract Deep learning has transformed computer vision for precision agriculture, yet apple orchard monitoring remains limited by dataset constraints. The lack of diverse, realistic datasets and the difficulty of annotating dense, heterogeneous scenes. Existing datasets overlook different growth stages and stereo imagery, both essential for realistic 3D modeling of orchards and tasks like fruit localization, yield estimation, and structural analysis. To address these gaps, we present AppleGrowthVision, a large-scale dataset comprising two subsets. The first includes 9,317 high resolution stereo images collected from a farm in Brandenburg (Germany), covering six agriculturally validated growth stages over a full growth cycle. The second subset consists of 1,125 densely annotated images from the same farm in Brandenburg and one in Pillnitz (Germany), containing a total of 31,084 apple labels. AppleGrowthVision provides stereo-image data with agriculturally validated growth stages, enabling precise phenological analysis and 3D reconstructions. Extending MinneApple with our data improves YOLOv8 performance by 7.69 % in terms of F1-score, while adding it to MinneApple and MAD boosts Faster R-CNN F1-score by 31.06 %. Additionally, six BBCH stages were predicted with over 95 % accuracy using VGG16, ResNet152, DenseNet201, and MobileNetv2. AppleGrowthVision bridges the gap between agricultural science and computer vision, by enabling the development of robust models for fruit detection, growth modeling, and 3D analysis in precision agriculture. Future work includes improving annotation, enhancing 3D reconstruction, and extending multimodal analysis across all growth stages. 1 Introduction Agricultural digitalization demands sophisticated datasets to develop advanced decision support systems and automate labor intensive tasks such as fruit picking. Monitoring apple orchards presents unique computational challenges due to highly variable environmental conditions—fluctuating light, unpredictable weather patterns, and dramatic seasonal transformations—all of which significantly impact image quality and consistency. These constraints have severely limited the development of robust models capable of reliable crop management and surveillance across diverse conditions [ 18 , 26 , 9 ] . Large-scale, temporally diverse orchard imagery offers transformative potential for precision agriculture. Such datasets significantly enhance yield estimation by enabling computer vision models to learn intricate fruit development patterns, effectively manage complex occlusions between densely clustered apples, and generate statistically reliable predictions, directly addressing persistent challenges in resource allocation and post-harvest waste [ 11 , 37 , 15 ] . Furthermore, comprehensive visual data improves model adaptability to dynamic field conditions, substantially enhancing algorithmic robustness and cross-orchard generalization while minimizing costly recalibration requirements [ 21 ] . We introduce AppleGrowthVision, the first publicly available dataset capturing apple orchards throughout a complete phenological cycle with calibrated stereo imagery. The dataset is publicly available at fraunhoferhhi.github.io/AppleGrowthVision . Our dataset uniquely provides scientifically validated agricultural annotations according to the ”Biologische Bundesanstalt, Bundessortenamt, und Chemische Industrie” (BBCH) scale [ 10 ] , documenting six critical growth stages of apple trees. The entire Brandenburg collection comprises precisely calibrated stereo image pairs, enabling accurate three-dimensional analysis and reconstruction. Our contributions include: • A comprehensive, multi-site dataset integrating orchards from different regions and complementary existing datasets, demonstrating quantifiable improvements of up to 28 % in apple detection accuracy and count estimation. • State-of-the-art classification models fine-tuned specifically for precise BBCH growth stage prediction in apple trees, achieving over 95% accuracy across growth stages and enabling data-driven phenological monitoring and intervention planning. • Novel computational techniques for reconstructing complex, large-scale orchard environments in three dimensions, supporting structural analysis, pruning optimization, and precisely targeted agricultural interventions. 2 Related Work Fruit detection in orchard environments has advanced significantly with the several specialized apple tree datasets. The MinneApple dataset provides 1,000 high-resolution images with over 41,000 annotated apple instances using polygonal masks to enhance object detection, segmentation, and fruit counting in orchards. This dataset supports patch-based fruit clustering analysis to facilitate yield estimation [ 11 ] . The dataset from Washington State University consists of 2,298 images from robotic harvesting studies, featuring multiple apple varieties under different lighting conditions in modern fruiting wall architectures [ 1 ] . Michigan State University’s apple detection dataset comprises of 1,246 RGB images of apple trees with varied lighting and occlusion conditions, developed alongside the Occluder-Occludee Relational Network (O2RNet), for detecting clustered apples [ 3 ] . The recent Monastery Apple Dataset (MAD) introduced a semi-supervised dataset comprising 5,545 images collected via drone from 16 apple trees in a German monastery, including 105 labeled images to improve detection across diverse lighting conditions [ 15 ] . While these datasets have advanced fruit detection capabilities, they lack critical growth stage annotations and typically cover limited time frames within a single growing season. AppleGrowthVision adresses these limitations by capturing an entire growth cycle with expert-validated phenological stages, enabling comprehensive analysis of fruit development and detection under real-world conditions. Phenology serves as a fundamental metric for plant classification and comparison. The extended BBCH scale represents a standardized system for describing vegetative stages in both research and agricultural management. Initially introduced by Hacke et al. in 1992 [ 10 ] and later extended for apple trees by Meier in 1994 [ 36 ] , this scale has been adopted for numerous plant species [ 22 ] . The BBCH system employs a two-digit coding system ranging from 0 to 99, where the first digit indicates the principal growth stage (Table 1 ), and the second digit defines secondary stages. This system’s key advantage is its consistent representation of similar phenological stages across diverse plant species. Recent work by Nguyen et al. [ 25 ] focused on annotating apple tree data for two principal growth stages and fine-tuneing multiple network architectures for classification. BBCH first digit principal stage 0 Bud development 1 Leaf development 3 Shoot development 5 Inflorescence emergence 6 Flowering 7 Development of fruit 8 Maturity of fruit and seed 9 Senescence, beginning of dormancy Table 1 : Used BBCH Principal growth stages for pome fruit [ 36 ] . Three-dimensional reconstruction of orchard environments presents substantial technical challenges. Current approaches typically operate under controlled laboratory conditions to eliminate wind effects and scene complexity [ 8 , 4 ] , or focus exclusively on leafless trees to simplify the reconstruction task [ 39 ] . CherryPicker [ 23 ] require up to 250 images per tree making detailed reconstruction of extensive orchards computationally prohibitive and operationally impractical. In contrast, AppleGrowthVision demonstrates that using a calibrated stereo camera setup dramatically improves reconstruction efficiency, enabling detailed 3D modeling of large orchard scenes with just 18 images per tree. This approach makes comprehensive digital twins of commercial orchards feasible for the first time, supporting advanced structural analysis and precision agricultural interventions at scale. 3 The AppleGrowthVision Dataset In response to these limitations, we introduce the AppleGrowthVision Dataset. A comprehensive, longitudinal resource that captures one growth cycle of apple trees. Unlike previous datasets, AppleGrowthVision includes high-resolution images for robust fruit detection and integrates expert-validated phenological data. This integration provides a dynamic view of apple development across different stages, bridging the gap between static image collections and the temporal complexity of real-world orchard environments. By allowing analyses that account for seasonal changes