Safety is paramount for mobile robotic platforms such as self-driving cars and unmanned aerial vehicles. This work is devoted to a task that is indispensable for safety yet was largely overlooked in the past - detecting obstacles that are of very thin structures, such as wires, cables and tree branches. This is a challenging problem, as thin objects can be problematic for active sensors such as lidar and sonar and even for stereo cameras. In this work, we propose to use video sequences for thin obstacle detection. We represent obstacles with edges in the video frames, and reconstruct them in 3D using efficient edge-based visual odometry techniques. We provide both a monocular camera solution and a stereo camera solution. The former incorporates Inertial Measurement Unit (IMU) data to solve scale ambiguity, while the latter enjoys a novel, purely vision-based solution. Experiments demonstrated that the proposed methods are fast and able to detect thin obstacles robustly and accurately under various conditions.

Background: High myopia (HM) is a major risk factor for glaucoma. However, glaucomatous optic neuropathy is often undiagnosed owing to atypical structural alterations with axial elongation. Moreover, an algorithm to detect glaucoma in highly myopic eyes has not yet been reported.Methods: We recruited 2643 colour fundus photographs to train a ResNet‐50 network for discriminating eyes with highly myopic glaucoma (HMG) from HM or glaucoma alone. We employed a 10‐fold cross‐validation strategy to evaluate the model's performance and applicability across diverse patient groups. Multiple metrics were computed to gauge the model's diagnostic process. The diagnostic ability of the model was then juxtaposed with those made by ophthalmologists to determine concordance. The gradient‐weighted class activation maps were used for visual explanations.Results: Our model demonstrated an overall accuracy of 97.7% with an area under the curve of 98.6% (sensitivity, 91.2%; specificity, 98.0%) for the differential diagnosis among HM, glaucoma, HMG and normal controls. These metrics notably outperformed the diagnostic performances of two attending ophthalmologists, who achieved accuracies of 64.7% and 69.9%. The activation maps derived from the model suggested that the most discriminative lesions for diagnosing HMG were predominantly in the disc, peripapillary area and inferior region of the disc, which are often displayed with a tessellated fundus. These results were slightly different from the understanding of the attending ophthalmologists.Conclusions: Our proposed model demonstrates high efficacy and suggests specific features for distinguishing eyes with HMG, enabling potential clinical value in assisting the intricate diagnosis of this vision‐threatening disease..

Despite recent progress, reconstructing outdoor scenes in 3D from movable platforms remains a highly difficult endeavour. Challenges include low frame rates, occlusions, large distortions and difficult lighting conditions. In this paper, we leverage the fact that the larger the reconstructed area, the more likely objects of similar type and shape will occur in the scene. This is particularly true for outdoor scenes where buildings and vehicles often suffer from missing texture or reflections, but share similarity in 3D shape. We take advantage of this shape similarity by localizing objects using detectors and jointly reconstructing them while learning a volumetric model of their shape. This allows us to reduce noise while completing missing surfaces as objects of similar shape benefit from all observations for the respective category. We evaluate our approach with respect to LIDAR ground truth on a novel challenging suburban dataset and show its advantages over the state-of-the-art.

Few-shot Learning (FSL) which aims to learn from few labeled training data is becoming a popular research topic, due to the expensive labeling cost in many real-world applications. One kind of successful FSL method learns to compare the testing (query) image and training (support) image by simply concatenating the features of two images and feeding it into the neural network. However, with few labeled data in each class, the neural network has difficulty in learning or comparing the local features of two images. Such simple image-level comparison may cause serious mis-classification. To solve this problem, we propose Augmented Bi-path Network (ABNet) for learning to compare both global and local features on multi-scales. Specifically, the salient patches are extracted and embedded as the local features for every image. Then, the model learns to augment the features for better robustness. Finally, the model learns to compare global and local features separately, i.e., in two paths, before merging the similarities. Extensive experiments show that the proposed ABNet outperforms the state-of-the-art methods. Both quantitative and visual ablation studies are provided to verify that the proposed modules lead to more precise comparison results.

Background: The clinical translation of AI-assisted surgical navigation and digital twin modeling is hindered by the absence of a standardized, reusable anatomical spatial reference framework, as current one-dimensional or two-dimensional localization methods for the Segond fracture bed inadequately represent three-dimensional bony spatial relationships for intelligent navigation systems. Objective: To meet the needs of AI- and digital twin–driven precision orthopedics, this study constructs a CT-based three-dimensional coordinate system of the knee using tibial landmarks, applies a normalization algorithm for standardized Segond fracture localization, and validates a reusable anatomical reference model, thereby bridging the translational gap in AI localization and offering a standardized computational framework for knee surgical planning and digital twin modeling. Methods: A retrospective review was conducted on 30 patients with left-sided Segond fracture who underwent thin-slice knee CT examinations at Beijing Shunyi Hospital between June 2010 and December 2025. Patients were randomly assigned to a modeling group ( n = 15) and an independent validation group ( n = 15). Using 3D Slicer software, the following landmarks were annotated: the most lateral point of the fibular head (P1), the midpoint of the lateral edge of the tibial tubercle (P2), the most superior point of Gerdy's tubercle (P3), and the center of the Segond fracture bed (P4). A right-handed orthogonal local coordinate system was constructed with P2 as the origin, based on P1, P2, and P3, and scale normalization was performed using the distance from P2 to P1. The modeling group was used to estimate the preliminary single center anatomical prior of the fracture bed in the normalized tibial local coordinate system(N-TLS). In the validation group, only the native global RAS coordinates of P1, P2, and P3 were input, and the predicted native global RAS coordinates of P4 were obtained via inverse geometric mapping. Localization accuracy was evaluated using the three-dimensional Euclidean distance between the predicted coordinates and the CT-measured reference coordinates. Results: The mean normalized coordinates of the Segond fracture bed center in the modeling group were (0.8209, 0.4300, −0.0078). Based on this anatomical prior, the deterministic geometric estimator achieved a mean localization error of 3.21 mm in the 15-case single-center independent validation cohort, with a standard deviation of 1.23 mm, a median error of 2.75 mm, a maximum error of 5.33 mm, and a minimum error of 1.68 mm. These findings indicate that the Segond fracture bed center exhibits a population-level distribution pattern within this local coordinate framework that is amenable to geometric localization. Conclusion: This study established a novel three-dimensional CT-based computational coordinate system for the knee joint using tibial bony landmarks, characterized the standardized three-dimensional spatial distribution pattern of the center of the Segond fracture bed, and developed a geometric prediction algorithm model for the fracture center point. The model demonstrated favorable millimeter-level localization accuracy, providing a standardized computational anatomical reference for digital surgical planning and future knee digital twin modeling, and offering methodological support for evidence-based clinical decision-making and intelligent orthopedic healthcare systems.

Automated anatomical labeling plays a crucial part for computer-assisted diagnostic systems targeting coronary artery diseases. Established from medical practice, the labeling conventions intrinsically carry profound prior knowledge about the results, indicating what outcome is favorable and what is illegal. However, the prior has been largely neglected by existing works. Drawing inspiration from syntax parsing in the NLP domain, we propose a neural stochastic grammar parser for anatomical labeling. Our method captures the essential parental and sibling dependencies between vessel segments, incorporates structural prior in a principled and interpretable manner, while retaining the learning capabilities of deep models. Experiments show encouraging results both for the robustness and accuracy of our method.