3D semantic scene understanding tasks have achieved great success with the emergence of deep learning, but often require a huge amount of manually annotated training data. To alleviate the annotation cost, we propose the first weakly-supervised 3D instance segmentation method that only requires categorical semantic labels as supervision, and we do not need instance-level labels. The required semantic annotations can be either dense or extreme sparse (e.g. 0.02% of total points). Even without having any instance-related ground-truth, we design an approach to break point clouds into raw fragments and find the most confident samples for learning instance centroids. Furthermore, we construct a recomposed dataset using pseudo instances, which is used to learn our defined multilevel shape-aware objectness signal. An asymmetrical object inference algorithm is followed to process core points and boundary points with different strategies, and generate high-quality pseudo instance labels to guide iterative training. Experiments demonstrate that our method can achieve comparable results with recent fully supervised methods. By generating pseudo instance labels from categorical semantic labels, our designed approach can also assist existing methods for learning 3D instance segmentation at reduced annotation cost.
Optimal control of obstacle problems arises in a wide range of applications and is computationally challenging due to the nonsmoothness, nonlinearity, and bilevel structure of the underlying problems. Classical numerical methods typically rely on mesh-based discretizations and require solving a sequence of computationally expensive subproblems, while deep learning approaches for this class of problems remain largely unexplored. In this work, we propose a bilevel deep learning framework whose core training algorithm is the computationally efficient Single-Loop Stochastic First-Order Bilevel Algorithm (S2-FOBA), which avoids nested optimization, uses only first-order information, and does not rely on restrictive lower-level uniqueness assumptions. The proposed method is mesh-free, scalable to high-dimensional settings and complex domains, and avoids the repeated solution of discretized subproblems. It employs constraint-embedding neural networks to approximate both the state and the control, thereby enforcing admissibility by construction while preserving the intrinsic bilevel structure of the problem. We further analyze the convergence behavior of S2-FOBA under mild conditions. Numerical experiments on benchmark problems, including two-dimensional examples with regular and irregular obstacles, a five-dimensional example, and a problem posed on a complex domain, demonstrate the effectiveness, efficiency, and robust constraint enforcement of the proposed approach.
In the emerging field of urban digital twins (UDTs), advancing intelligent road inspection (IRI) vehicles with automatic road crack detection systems is essential for maintaining civil infrastructure. Over the past decade, deep learning-based road crack detection methods have been developed to detect cracks more efficiently, accurately, and objectively, with the goal of replacing manual visual inspection. Nonetheless, comprehensive comparative analyses of state-of-the-art (SoTA) deep learning techniques, especially data-fusion and label-efficient algorithms for this task, remain limited. This paper thoroughly reviews the SoTA deep learning-based algorithms, including (1) supervised, (2) unsupervised, (3) semi-supervised, and (4) weakly-supervised methods developed for road crack detection. In addition, we establish UDTIRI-Crack, a curated and standardized benchmark suite comprising 2500 images from seven public annotated datasets, and develop an online benchmark platform to provide standardized, reproducible, and extensible evaluation of algorithms. Comprehensive experiments are conducted to compare the detection performance, computational efficiency, and generalizability of public SoTA deep learning-based algorithms for road crack detection. The feasibility of foundation models and vision-language foundation models for road crack detection is further explored. Finally, the existing challenges and future development trends of deep learning-based road crack detection algorithms are discussed. We believe this review, together with the released dataset and platform, can serve as practical guidance for developing intelligent road detection vehicles with the next-generation road condition assessment systems. The benchmark UDTIRI-Crack is available at https://udtiri.com/submission/.
Accurate fuel consumption prediction is often limited by data scarcity, especially as current state-of-the-art learning-based approaches are typically data-hungry. In practice, historical records are unevenly distributed across the fleet. While a few well-equipped vessels provide dense, high-frequency data, most vessels have only sparse records due to short observation periods, manual log collection, or maintenance-related data gaps. Consequently, training robust models for vessels with limited historical data is challenging. To address this challenge, we propose F2V-Former, a novel fleet-to-vessel transfer learning framework that leverages abundant data collected across the fleet to pre-train a unified model before fine-tuning it for target vessels. To improve adaptability to heterogeneous vessels with different capacities, F2V-Former integrates static vessel attributes with dynamic operational sequences via static-guided conditional layer normalization and cross-attention. For parameter-efficient vessel-specific adaptation in few-shot settings, we adopt low-rank adaptation (LoRA) to fine-tune the network. Experiments on a real-world container shipping fleet show that F2V-Former achieves the best performance in zero-shot and in most few-shot settings. Notably, in zero-shot experiments, F2V-Former achieves a MAPE of 14% without vessel-specific fine-tuning, demonstrating its reliable generalization among the new vessels in the fleet. Further experiments, including error evolution analysis, sensitivity analysis, and ablation studies, are conducted to validate the effectiveness of the proposed framework. These results demonstrate that F2V-Former achieves high prediction accuracy and reliable cross-vessel generalization in data-scarce scenarios.
Large Language Models (LLMs) are rapidly transforming intelligent healthcare and medicine. However, their successful clinical deployment depends on reliable and explainable medical reasoning rather than surface-level text generation. This survey reviews LLM reasoning in medicine. It covers foundational datasets, training-time optimization, test-time reasoning, collaborative agents, and reasoning evaluation. We distinguish two complementary design axes. Individual-model optimization improves the reasoning behavior of a model or module through supervised fine-tuning, reinforcement-based fine-tuning, or test-time control. Collaborative-agent design specifies how one or more such modules exchange evidence, invoke tools, and aggregate decisions at the system level. Collaborative-agent design does not replace individual-model optimization. Contemporary agent systems commonly compose prompted or individually trained modules, including SFT- and GRPO-optimized planners, routers, and aggregators. This compositional view clarifies both the promise and the current limitations of agents. Structured collaboration can support transparency, error recovery, and heterogeneous evidence integration in workflow-oriented tasks. However, it does not provide an inherent advantage over a strong single LLM on narrowly scoped questions. Combining individual-model optimization with collaborative-agent design offers substantial potential for advancing future medical intelligence.
