26 June 2024 in IEEE Transactions on Pattern Analysis and Machine Intelligence
The production of food, feed, fiber, and fuel is a key task of agriculture, which has to cope with many challenges in the upcoming decades, e.g., a higher demand, climate change, lack of workers, and the availability of arable land. Vision systems can support making better and more sustainable field management decisions, but also support the breeding of new crop varieties by allowing temporally dense and reproducible measurements. Recently, agricultural robotics got an increasing interest in the vision and robotics communities since it is a promising avenue for coping with the aforementioned lack of workers and enabling more sustainable production. While large datasets and benchmarks in other domains are readily available and enable significant progress, agricultural datasets and benchmarks are comparably rare. We present an annotated dataset and benchmarks for the semantic interpretation of real agricultural fields. Our dataset recorded with a UAV provides high-quality, pixel-wise annotations of crops and weeds, but also crop leaf instances at the same time. Furthermore, we provide benchmarks for various tasks on a hidden test set comprised of different fields: known fields covered by the training data and a completely unseen field.
20 June 2022 in IEEE Robotics and Automation Letters
Ego-motion estimation is a fundamental building block of any autonomous system that needs to navigate in an environment. In large-scale outdoor scenes, 3D LiDARs are often used for this task, as they provide a large number of range measurements at high precision. In this paper, we propose a novel approach that exploits the intensity channel of 3D LiDAR scans to compute an accurate odometry estimate at a high frequency. In contrast to existing methods that operate on full point clouds, our approach extracts a sparse set of salient points from intensity images using data-driven feature extraction architectures originally designed for RGB images. These salient points are then used to compute the relative pose between successive scans. Furthermore, we propose a novel self-supervised procedure to fine-tune the feature extraction network online during navigation, which exploits the estimated relative motion but does not require ground truth data. The experimental evaluation suggests that the proposed approach provides a solid ego-motion estimation at a much higher frequency than the sensor frame rate while improving its estimation accuracy online.
21 June 2023 in IEEE Robotics and Automation Letters
Measuring plant traits with high throughput allows breeders to monitor and select the best cultivars for subsequent breeding generations. This can enable farmers to improve yield to produce more food, feed, and fiber. Current breeding practices involve extracting leaf parameters on a small subset of the leaves present in the breeding plots, while still requiring substantial manual labor. To automate this process, an important step is the precise distinction between separate leaves, which is the problem we address in this letter. We exploit recent advancements in 3D deep learning to build a convolutional neural network that learns to segment individual leaves. As done in current breeding practices, we select a subset of leaves to be used for phenotypic trait evaluation as this allows us to alleviate the influence of segmentation errors on the phenotypic trait estimation. To this extent we propose to use an additional neural network to predict the quality of each segmented leaf and discard inaccurate leaf instances. The experiments show that our network yields higher segmentation accuracy on sugar beet breeding plots planted under the supervision of the German Federal Office for Plant Varieties. Furthermore, we show that our neural network helps in filtering out leaves with lower segmentation accuracy.
4 August 2025 in IEEE Robotics and Automation Letters
Automatic fruit growth monitoring plays a vital role in advancing precision agriculture. Tracking the evolution of fruits over time is essential to monitor their development and optimize production. The ability to recognize fruits over periods of time, even with drastic scene changes, is a required capability of agricultural robots. This letter presents a system that allows long-term fruit tracking in 3D data. It generates instance-segmented 3D representations of plants at various growth stages over time, utilizing only consumer-grade RGB-D cameras installed on a mobile robot. Our approach first performs instance segmentation on each image in a sequence. Then, by exploiting geometric information and depth maps, we track the same instances throughout the sequence. We produce a 3D point cloud containing instances, exploiting odometry information and 3D semantic mapping. Once our robot performs a new recording at a different plant growth stage, it associates each fruit with the previously built 3D cloud and update the model. We validate the system in a real-world glasshouse environment in Bonn, Germany. Experimental results demonstrate that our system outperforms existing baselines even though it relies only on annotated images and operates at frame-rate, allowing the deployment on a real robot.
11 November 2026 in The International Journal of Robotics Research
Robots need to be able to understand their surroundings in order to operate safely and robustly, and to interact with the surrounding environment. Robots deployed in unconstrained real-world scenarios must additionally be able to deal with novel situations and objects that have never been seen before. In this article, we tackle the problem of open-world panoptic segmentation, that is, the task of discovering new semantic categories and new object instances at test time, while enforcing consistency among the categories that we incrementally discover. We present Con2MAV, a general method for open-world panoptic segmentation. Experiments across a wide range of datasets, from road scenes to underwater environments, highlight its compelling capabilities in open-world segmentation and its competitive performance on known classes. We will open-source the implementation of our approach upon acceptance. In addition, we propose PANIC ( P anoptic AN omalies I n C ontext), a benchmark for evaluating open-world segmentation tasks in autonomous driving scenarios. This dataset, recorded with a multi-modal sensor suite mounted on a car, and then manually annotated, provides high-quality, pixel-wise annotations of anomalous objects at both semantic and instance level. PANIC contains 800 images, more than 50 unknown classes, that is, classes that do not appear in the training set, and over 4000 object instances, providing a comprehensive benchmark for evaluating open-world segmentation methods in autonomous driving scenarios. We provide competitions for multiple open-world segmentation tasks on a hidden test set. Our dataset and competitions are available at https://www.ipb.uni-bonn.de/data/panic .