Multi-task visual grounding (MTVG) covers two sub-tasks: Referring Expression Comprehension (REC) and Referring Expression Segmentation (RES). Existing approaches typically follow a pipeline of independent visual and linguistic feature extraction, a cross-modal interaction module, and separate prediction heads. Despite this progress, this line has two limitations: (1) lacking a careful design of where to inject language into the visual backbone, they need an extra cross-modal fusion module, increasing model complexity and computational cost; (2) the relationship between REC and RES is underexploited, although the two tasks target the same object and share the same center. To address these issues, we propose a Progressive Language-guided Visual Learning framework (PLVL), which mines the inherent visual features while progressively injecting language information to learn linguistic-related visual features. Thus PLVL fully leverages language guidance without an additional cross-modal fusion module. Moreover, since RES is a pixel-wise classification task and both tasks target the same object, their center cues may be correlated. We therefore adopt a shared-representation multi-task head with grouped feature routing while keeping separate task-specific losses. Experiments on RefCOCO validate our language-guided backbone and shared-representation head for both REC and RES. On RefCOCOg val(U), PLVL surpasses EEVG by 1.98% (REC) and 1.06% (RES) under traditional training, and by 4.01% (REC) and 3.69% (RES) with multi-dataset pre-training. Code is available at https://github.com/jcwang0602/PLVL
Field emission coupled with molecular dynamics simulation (FEcMD) software package is a computational tool for studying atomic structure evolution, structural deformation, phase transitions, recrystallization as well as electron emission characteristics of micro- and nano-protrusions and nanowires consisting of elemental metals or multi-component alloys by means of multi-physics and multi-scale methodology. Current implementations of molecular dynamics simulation coupled with multi-scale electrodynamics (ED) and heat conduction (HC) in FEcMD program are advanced mainly in the two aspects as follows. In electrodynamics, the FEcMD program incorporates the space charge interactions (space charge potential and exchange-correlation effects) in the self-consistent solved Poisson-Schrödinger equation with Wentzel-Kramers-Brillouin-Jeffreys (WKBJ) approximation to evaluate the field emission current density and the related resistive heating process more reliably for nanowires or nano-protrusions especially for nano-gaps between two metal electrodes. Meanwhile, the two-temperature heat conduction model is implemented in electrodynamics coupled with molecular dynamics simulations (ED-MD), providing more dedicated descriptions for the hierarchical electron-phonon two-channel heat conduction mechanism and the temperature evolutions of electron and phonon subsystems under the radiofrequency (RF) or pulse electric fields. Benchmark tests are performed for some key implementations in FEcMD software to validate the numerical results, and also to demonstrate the use of program to study the atomic structure evolution of metal nano-structures under electric field and heating processes.
This paper presents a numerical approach to the stochastic obstacle problem using the stochastic Galerkin (SG) method. Due to the low regularity of the solution, linear finite elements are employed in both the physical and random variable spaces. Properties of random fields and variational inequalities of the first kind are employed to establish the well-posedness of the problem. Finite element spaces are introduced to construct suitable approximation subspaces, and a comprehensive SG formulation is proposed to solve the stochastic obstacle problem. Well-posedness of the discrete formulation is shown and an optimal error estimate for the numerical solution in the H1-norm is derived. Numerical experiments validate the effectiveness of the SG method, showing that the expectation error converges at a rate of O(h) in the H1-norm, consistent with theoretical predictions.
Recognition of cross-border trade documents degrades substantially when images are of low quality, suffering from scanning creases, stains, and character adhesion—issues that are especially severe for non-Latin scripts such as Russian. To address these challenges, we propose an ACmix (Attention–Convolution Mixture) dynamic feature fusion architecture integrated into a lightweight OCR framework. The design combines a parallel self-attention pathway to model global semantic dependencies with a convolutional branch to capture local texture features, forming a global–local dual-path interaction that improves discrimination for blurred and adhered characters. We integrate ACmix modules into a PPOCRv3 backbone and construct the ACmix-PPOCRv3 model. Experiments on ICDAR-2015 (English), RDIOD (Russian), and a mixed-language dataset show improvements: English accuracy reaches 92.53
In human-centric manufacturing, collaborative robots (cobots) are increasingly expected to continually acquire knowledge as products and workflows evolve, and learning from human demonstration videos has emerged as a flexible means to this end. Realizing this paradigm requires perception that can be deployed with minimal hardware overhead and that generates executable task knowledge for robotic manipulation. Existing video-based methods rarely satisfy both requirements. Geometric pipelines rely on CAD models or multi-camera rigs, hindering scalable deployment across diverse workpieces, while pre-trained vision-language models require demonstration corpora rarely available in high-mix, low-volume manufacturing. In addition, existing methods recover only low-level geometry or high-level semantic labels in isolation, leaving a gap between geometric perception and executable task knowledge. To address these limitations, we propose Video2Knowledge, a video-based learning framework that takes a single monocular RGB video as input and distills it into temporally consistent task knowledge, covering action sequences, object state transitions, and spatial relations, without requiring additional sensors, manual CAD models, or task-specific retraining. The framework couples a physics-aware 6D object pose corrector with hand-object interaction analysis, and consolidates the results into a structured knowledge representation that maps directly to robot action sequences. Experimental evaluation on the HA-ViD and Assembly101 datasets demonstrates that the physics-aware correction yields improved pose coherence, while the extracted task knowledge faithfully recovers demonstrated action semantics and relative geometric configurations. A real-world case study further validates the transfer of task knowledge from a monocular demonstration video to physical cobot execution, confirming the practical deployability in manufacturing scenarios.
