From Occlusion to 3D: Amodal Completion Enables Single-View Wheat Reconstruction

Abstract

Abstract Occlusion is a major factor limiting accurate three-dimensional (3D) wheat phenotyping. In natural growth conditions, overlapping spikes, leaves, and stems often make only partial target regions visible in single-view images, hindering complete and reliable 3D reconstruction. To address this problem, this study proposes an amodal completion-driven framework for single-view 3D reconstruction of occluded wheat. The framework first uses visible prompts to recover the complete appearance and structural cues of occluded targets, and then feeds the completed images into single-view 3D reconstruction models to generate complete 3D structures. To support model training and evaluation, we construct the MMWO (Multi-view Multi-instance Wheat Occlusion) dataset from MMW by synthesizing diverse occlusion samples through organ-level cutouts, random geometric transformations, and region-constrained pasting, with annotations including visible masks, occlusion masks, and complete target images. Six representative reconstruction methods, including Direct3D, Real3D, SF3D, Spar3D, TRELLIS.2, and Hunyuan3D, are systematically evaluated. Hunyuan3D achieves the best geometric performance, with the lowest mean CD-L 1 and CD-L 2 values of 0.1286 and 0.0536, and the highest mean F-score of 0.5668. SF3D achieves the best rendering quality in terms of PSNR, SSIM, and LPIPS. In addition, Pix2Gestalt completion reduces the estimation errors of spike length, width, and area from 9.31%, 10.89%, and 32.23% to 4.64%, 9.70%, and 9.45%, respectively. These results demonstrate that amodal completion can effectively alleviate occlusion-induced information loss and provide more complete structural priors for single-view 3D wheat reconstruction. This study offers a feasible solution for robust 3D wheat phenotyping under occlusion and provides a systematic reference for applying general-purpose 3D generative models to agricultural phenotyping.

Plant phenotyping relevance

遮蔽下小麦の単視点3D再構成とアモーダル補完を開発し、データセット構築、複数手法の系統評価、器官形質推定誤差の検証まで行っており、植物フェノタイピング手法が中心である。

Code and data availability

The paper introduces the MMWO wheat occlusion dataset (built from MMW), a fine-tuned Pix2Gestalt model, and an evaluation pipeline, but the supplied blocks contain no public deposit, repository URL, or availability statement for the dataset, code, or trained model. Supplementary files are listed only as filenames with

No evidence-backed public reproduction asset is currently recorded.