Iterative Online 3D Reconstruction from RGB Images.

3D reconstruction is the computer vision task of reconstructing the 3D shape of an object from multiple 2D images. Most existing algorithms for this task are designed for offline settings, producing a single reconstruction from a batch of images taken from diverse viewpoints. Alongside reconstruction accuracy, additional considerations arise when 3D reconstructions are used in real-time processing pipelines for applications such as robot navigation or manipulation. In these cases, an accurate 3D reconstruction is already required while the data gathering is still in progress. In this paper, we demonstrate how existing batch-based reconstruction algorithms lead to suboptimal reconstruction quality when used for online, iterative 3D reconstruction and propose appropriate modifications to the existing Pix2Vox++ architecture. When additional viewpoints become available at a high rate, e.g., from a camera mounted on a drone, selecting the most informative viewpoints is important in order to mitigate long term memory loss and to reduce the computational footprint. We present qualitative and quantitative results on the optimal selection of viewpoints and show that state-of-the-art reconstruction quality is already obtained with elementary selection algorithms.

Ultra-high-resolution UAV-imaging and supervised deep learning for accurate detection of Alternaria solani in potato fields.

Alternaria solani is the second most devastating foliar pathogen of potato crops worldwide, causing premature defoliation of the plants. This disease is currently prevented through the regular application of detrimental crop protection products and is guided by early warnings based on weather predictions and visual observations by farmers. To reduce the use of crop protection products, without additional production losses, it would be beneficial to be able to automatically detect Alternaria solani in potato fields. In recent years, the potential of deep learning in precision agriculture is receiving increasing research attention. Convolutional Neural Networks (CNNs) are currently the state of the art, but also come with challenges, especially regarding in-field robustness. This stems from the fact that they are often trained on datasets that are limited in size or have been recorded in controlled environments, not necessarily representative of real-world settings. We collected a dataset consisting of ultra-high-resolution modified RGB UAV-imagery of both symptomatic and non-symptomatic potato crops in the field during various years and disease stages to cover the great variability in agricultural data. We developed a convolutional neural network to perform in-field detection of Alternaria, defined as a binary classification problem. Our model achieves a similar accuracy as several state-of-the-art models for disease detection, but has a much lower inference time, which enhances its practical applicability. By using training data of three consecutive growing seasons (2019, 2020 and 2021) and test data of an independent fourth year (2022), an F1 score of 0.93 is achieved. Furthermore, we evaluate how different properties of the dataset such as its size and class imbalance impact the obtained accuracy.