Answers are grounded in this paper and cite their sources.
In precision agriculture and intelligent breeding, high-quality, close-ground panoramic images of farmland are essential for large-scale, continuous monitoring of crop growth. However, a single near-ground image has a limited field of view, making it difficult to assess crop growth comprehensively; by contrast, panoramic images can integrate information from multiple images to provide comprehensive coverage of farmland, thereby supporting precision farming and intelligent breeding. Although gantry-based acquisition is widely used in phenotypic research, near-ground photography is prone to issues such as similar textures, significant parallax, and variable lighting conditions, which can readily lead to ghosting, structural distortion, or even failure during image stitching. To address these issues, the paper proposes a near-ground field stitching algorithm based on spatial layout priors and crop object preservation, named as NFStitch. Specifically, first, camera translation prior is utilised to optimise feature matching, focusing on stable ground areas; secondly, a similar transformation matrix is constructed with the constraints as much as possible translation, and the optimization solution of the shooting trajectory protection items is integrated to realize the global alignment and protection of the farmland structure; finally, plant masks are obtained via SAM2, and their segmentation cost is increased in the energy function to guide the stitching lines away from crops. The experiment was validated on 22 datasets covering different growth stages of wheat. Compared to mainstream methods, this algorithm better balances geometric structure and crop integrity, providing an effective solution for generating high-quality panoramic images of farmland.
No takes yet. Share an insight, caveat, or question.
Zhou et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: