% IMPORTANT: The following is UTF-8 encoded.  This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.

@INPROCEEDINGS{Cherepashkin:1022310,
      author       = {Cherepashkin, Vsevolod and Yildiz, Erenus and Fischbach,
                      Andreas and Kobbelt, Leif and Scharr, Hanno},
      title        = {{D}eep {L}earning {B}ased 3d {R}econstruction for
                      {P}henotyping of {W}heat {S}eeds: a {D}ataset, {C}hallenge,
                      and {B}aseline {M}ethod},
      reportid     = {FZJ-2024-01428},
      pages        = {561-571},
      year         = {2023},
      abstract     = {We present a new data set for 3d wheat seed reconstruction,
                      propose a challenge, and provide baseline methods.
                      Individual plant seed properties influence early development
                      of plants and are thus of interest in plant phenotyping
                      experiments. Seed shape can be measured reliably from images
                      using volume carving, as done in robotic setups such as
                      phenoSeeder. However, about 36 images are needed to obtain a
                      suitably accurate 3d model, where image acquisition takes
                      approximately 20 s. For large-scale experiments with
                      thousands of seeds higher throughput is required limiting
                      image acquisition time. We present a deep-learning model
                      that reconstructs an approximate 3d point cloud from fewer
                      images, even only a single view. It has a significantly
                      lower error than linear regression, which has been actively
                      used so far in similar tasks. Using three images reduces
                      imaging time by a factor of 10, where relative errors of
                      volume length, width, and height are all around $2\%.$
                      Inference time from the neural network is negligibly short
                      compared with imaging time which enables this method for
                      real-time measurements and sorting.},
      month         = {Oct},
      date          = {2023-10-02},
      organization  = {ICCV 2023, Paris (France), 2 Oct 2023
                       - 6 Oct 2023},
      cin          = {IAS-8},
      cid          = {I:(DE-Juel1)IAS-8-20210421},
      pnm          = {5112 - Cross-Domain Algorithms, Tools, Methods Labs (ATMLs)
                      and Research Groups (POF4-511)},
      pid          = {G:(DE-HGF)POF4-5112},
      typ          = {PUB:(DE-HGF)8},
      doi          = {10.34734/FZJ-2024-01428},
      url          = {https://juser.fz-juelich.de/record/1022310},
}