Japan Geoscience Union Meeting 2022

Presentation information

[J] Poster

M (Multidisciplinary and Interdisciplinary) » M-TT Technology & Techniques

[M-TT46] Introducing metaverse to agriculture. Are we ready?

Fri. Jun 3, 2022 11:00 AM - 1:00 PM Online Poster Zoom Room (35) (Ch.35)

convener:Seishi Ninomiya(Graduate School of Agriculture and Life Sciences, the University of Tokyo), convener:Yukihiro Takahashi(Department of Cosmosciences, Graduate School of Science, Hokkaido University), Chairperson:Seishi Ninomiya(Graduate School of Agriculture and Life Sciences, the University of Tokyo)

11:00 AM - 1:00 PM

[MTT46-P08] Improved soybean pod counting and localization with feature level considered

*Jiangsan Zhao1, Akito Kaga2, Masayuki Hirafuji1, Seishi Ninomiya1, Wei Guo1 (1.Graduate School of Agriculture and Life Sciences, The University of Tokyo, 2.Crop Gene Function Group, Institute of Crop Science, NARO)

Keywords:soybean pod, point based counting

Effective soybean pod counting tools can be used for its yield prediction directly in the field ahead of the final harvest. An integrated approach that can manage both counting and localization will be ideal for subsequent analysis. The traditional way of object counting through predicting bounding boxes or density map over instances is significantly limited by the labor-intensive and error-prone labeling procedure. In order to quantify the soybean pods directly rather than in a sequential manner, we propose P2PNet+, which is built upon the point-based framework for joint crowd counting and individual localization, P2PNet. Several strategies are considered to adjust the architecture and subsequent post processing to maximize the model performance in soybean pod counting and localization. First, based on advantages of either higher resolution or higher-level features, one module collects lower-level features for localization while another one extracts higher-level features for classification applied; the combined outcome of two modules are used to further improve the final performance of the model. Second, as edge information is important for both classification and detection tasks, attention module is embedded into the network to make full use of the edge information to improve the model performance. At last, with the above efforts, predicted points located close enough to reference points still cannot be fully eliminated and result in higher prediction errors. Subsequent thresholding method is applied to merge closely located predictions to improve the accuracy of soybean pod counting and localization. Through training the model on cropped images of soybean plants from one side and testing on images taken from the opposite side, the superiority of proposed network over the original P2PNet is validated.