Evaluation of Image-Based 3D Plant Reconstruction Methods
Buch in deiner Nähe kaufen
Beschreibung
In this work we investigated the possibility of generating 3D models for real and complex plants using only high resolution video and an image-based 3D reconstruction pipeline. We have specifically tested the performance of two consumer cameras, contrasting the reconstructed point-cloud models with ground-truth data acquired with a state-of-the-art scanner.
We have shown that it is possible to reconstruct 3D models from 4K resolution video, useful for phenotyping. The quality of these reconstructions for answering biological questions depends on the resolution required by the plant scientist. For point clouds with a resolution between 0.3-1 mm, our reconstructed models reach more than 50% of the ground truth in terms of a precision and completeness measure. For lower resolutions, below 1 mm, the quality of the reconstructions is above 85%, and for resolutions lower than 1 cm, the models are 100% as good as the ground-truth data.
Additionally, we have found that video offers several advantages over scanning methods. In particular, it does not interfere with the plant’s growth patterns and the data collection time is reduced from hours of scanning to 2-4 minutes video recording. According to this study, image-based methods are a cost-effective alternative for plant phenotyping, delivering results comparable to those of much more expensive methods. There are still challenges regarding computation time, but our results are promising and point to exciting challenges and future research in the field of plant phenotyping.
Autor*in
Pablo Andres Amador Rodriguez
Themen in »Evaluation of Image-Based 3D Plant Reconstruction Methods«
3D Reconstruction COLMAP Computer Vision F-Score Image-based Multi-View Stereo Plant Reconstruction Stereo Vision Structure from Motion Structured-Light