Grapes Ready for Picking: Investigating the Readiness of Precision Viticulture Datasets for Scene Reconstruction

Michał Jan Włodarczyk1,†, Katarzyna Wasielewska-Michniewska1,†, Przemysław Hołda1, Maria Ganzha1, Marcin Paprzycki1, Nico Heider2, Martin Schieck2,†, Bogdan Franczyk2,†, Tasuku Makabe3, Iori Yanokura3, Kei Okada3
1Systems Research Institute, Polish Academy of Sciences, Warsaw, Poland 2Leipzig University, Germany 3University of Tokyo, Japan †Equal Contribution
Code Submit a dataset

Abstract

Public datasets of vineyard imagery were collected for detection, segmentation and disease mapping, and they are now being proposed as a source for 3D scene reconstruction. The properties that reconstruction depends on, such as view overlap, parallax and frame spacing, are often omitted from their documentation, so it is unclear which of these datasets can be reconstructed. The aim of this study is to measure that for every publicly available vineyard dataset. A structured search identifies eighteen datasets, of which ten contain overlapping views. These ten are characterized under a schema of thirty-six documented and measured properties, and every sequence that registers is reconstructed with one fixed pipeline of structure from motion followed by radiance field reconstruction via 3D Gaussian Splatting, giving twenty-eight reconstructions on two held-out splits. Registration is decided by the capture rather than by the dataset, novel-view quality varies as much within a dataset as between datasets, and frame spacing is the one capture property that limits reconstruction. For the datasets that fail to register or that reconstruct with visible artifacts, three preprocessing steps are tested as ablations. Of these three steps, one restores registration, one improves rendered quality only marginally, and one, which masks static content that contaminates the reconstruction, exposes a fault in the camera motion that masking alone cannot fix. The schema and the pipeline give a procedure that can be applied to any vineyard dataset, existing or new, and the results are published as an open online catalog meant to help researchers choose data and to grow as further datasets are assessed in the same way.

The catalog

Found a value that differs from a dataset's documentation or data? Report a correction

Readiness

Readiness decision tree First question: does the imagery provide overlapping views? If yes, do at least half of the frames register in every repeated run? No leads to unstable (ripening). Yes leads to: does registration or color require a validated preprocessing step? Yes leads to ready with treatment (ripening), and no leads to ready (ripe). If the imagery has no overlapping views: does it ship geometry or imagery that can constrain a reconstruction? Yes leads to prior only (ripening), and no leads to not ready (unripe). yes no no yes yes no yes no Does the imagery provide overlapping views? (sequence continuity) Do at least half of the frames register in every repeated run? (registered share, stability) Does it ship geometry or imagery that can constrain a reconstruction? (auxiliary 3D) Does registration or color require a validated preprocessing step? Unstable (ripening) Prior only (ripening) Not ready (unripe) Ready with treatment (ripening) Ready (ripe)
  • Does the imagery provide overlapping views? (sequence continuity)
    • yesDo at least half of the frames register in every repeated run? (registered share, stability)
      • noUnstable (ripening)
      • yesDoes registration or color require a validated preprocessing step?
        • yesReady with treatment (ripening)
        • noReady (ripe)
    • noDoes it ship geometry or imagery that can constrain a reconstruction? (auxiliary 3D)
      • yesPrior only (ripening)
      • noNot ready (unripe)
Follow the questions above, from the top, to find the readiness class of a dataset.

Submit a dataset

This is a place for practitioners to share the vineyard data they collect, in academia and outside it. The ten datasets above give a baseline to compare against, and the paper gives a set of recommendations to consider when collecting data.

Open the submission form Submit by e-mail

Contact

Michał Jan Włodarczyk
michal.wlodarczyk@ibspan.waw.pl
corresponding author

Questions are also welcome in the repository issues.

Citation

@Article{s26216708,
AUTHOR = {Włodarczyk, Michał Jan and Wasielewska-Michniewska, Katarzyna and
          Hołda, Przemysław and Ganzha, Maria and Paprzycki, Marcin and
          Heider, Nico and Schieck, Martin and Franczyk, Bogdan and
          Makabe, Tasuku and Yanokura, Iori and Okada, Kei},
TITLE = {Grapes Ready for Picking: Investigating the Readiness of
         Precision Viticulture Datasets for Scene Reconstruction},
JOURNAL = {Sensors},
VOLUME = {26},
YEAR = {2026},
NUMBER = {21},
ARTICLE-NUMBER = {6708},
URL = {https://www.mdpi.com/1424-8220/26/21/6708},
ISSN = {1424-8220},
DOI = {10.3390/s26216708}
}