The advent of the 5G RedCap, the upcoming 6G and the proliferation of the Internet of Things (IoT) have catalyzed the rapid advancement of unmanned aerial vehicle (UAV) technology while also promoting UAVs' widespread application. In IoT-enabled environments where the global positioning system (GPS) signals are compromised, visual simultaneous localization and mapping (V-SLAM) technology has emerged as an effective positioning solution, valued for its reliability. However, the presence of dynamic elements in complex environments, such as pedestrians and vehicles, poses challenges to the positioning accuracy of UAVs employing V-SLAM for navigation. This paper proposes a dynamic feature filtering-based SLAM (DFF-SLAM) approach to eliminate the impact of dynamic factors in dynamic environments, thereby enhancing the positioning accuracy of UAVs in IoT-enabled complex environments. Firstly, a semantic detection thread is designed to identify semantic information in the scene and acquire prior dynamic targets, facilitating the filtering of prior dynamic feature points. Secondly, optical flow tracking conducted at each level of the image pyramid facilitates feature point matching across consecutive images. Finally, the epipolar geometry constraint is utilized to determine the motion status of remaining feature points, further filtering out dynamic feature points. Simulation results demonstrate that compared to traditional visual SLAM systems, the UAV equipped with the DFF-SLAM system achieves more accurate positioning and meets real-time positioning requirements when navigating through IoT enabled complex environments

Vision-Language Models have made significant progress on many perception-focused tasks.However, their progress on reasoningfocused tasks remains limited due to the lack of high-quality and diverse training data.In this work, we aim to address the scarcity of reasoning-focused multimodal datasets.We propose VisualWebInstruct, a novel approach that leverages search engines to create a diverse and high-quality dataset spanning multiple disciplines, including mathematics, physics, finance, and chemistry, etc. Starting with a meticulously selected set of 30,000 seed images, we employ Google Image Search to identify websites containing similar images.We collect and process HTML data from over 700K unique URLs.Through a pipeline of content extraction, filtering, and synthesis, we construct a dataset of approximately 900K question-answer (QA) pairs, with 40% consisting of visual QA pairs and the remaining comprising text-based QA pairs.Models finetuned on VisualWebInstruct demonstrate significant performance improvements: (1) finetuning on Llava-OV results in 10-20 absolute points improvement across benchmarks, and (2) fine-tuning from MAmmoTH-VL yields a 5 absolute points gain across benchmarks.Our best model, MAmmoTH-VL2, achieves the best known performance with SFT without RL within the 10B parameter class on MMMU-Pro (40.7),MathVerse (42.6), and DynaMath (55.7).These results highlight the effectiveness of our dataset in enhancing the reasoning capabilities of vision-language models for complex multimodal tasks.

Abstract The coasts of the Northeastern United States experience wind and flood damage as a result of extratropical cyclones (such as Nor’easters). However, recorded data is limited for hazard analysis and resilience evaluation. This paper describes a method that can efficiently augment the time series of extratropical cyclone severity measures (wind speed and flood elevation) by leveraging convolutional variational autoencoders, a deep learning technique. As a generative model, the proposed convolutional variational autoencoder learns the probability distribution of latent features (here, low-dimensional underlying characteristics of the time series) as a multivariate normal distribution. The augmented severity measure time series are then obtained in two steps: (1) randomly sampling data from the learned multivariate normal distribution and (2) using the randomly sampled data as the input to the decoder of the trained convolutional variational autoencoder to generate severity measures. A case study is conducted to evaluate the quality of the augmented wind speed and flood elevation time series of the extratropical cyclones recorded in Boston Harbor (adjacent to the city of Boston, Massachusetts). This study also demonstrates that the proposed method outperforms other established time series data augmentation methods (i.e., Dynamic Time Warping and Conditional Generative Adversarial Network). An application for hazard frequency analysis—modeling the joint probability of wind speed and flood elevation using a Gumbel copula—is also presented. The results demonstrate that the augmented data can significantly reduce the uncertainty of the estimated copula parameter while obtaining a goodness-of-fit metric value (negative log-likelihood) similar to that of the original, non-augmented data. The proposed convolutional variational autoencoder can be used to augment any time series data. The user-friendly codes are published.

Nonlinear time history analyses of structural models, widely used in civil engineering, can be time-consuming. For an urban scale, modeling a large number of structures and carrying out nonlinear time history analyses under different levels of hazard intensity is not practical. This can be resolved by the use of metamodeling, which can effectively reduce the computational cost while maintaining necessary engineering accuracy. Neural networks have been proven to be a powerful tool for metamodel development. However, training a reliable deep learning model requires a large training dataset that contains representative input-output relationships, which typically cannot be satisfied in practical applications. Practically, a trained deep learning model is a “black box”, which has limited (or no) generalizability. In this paper, a symbolic-based neural network that does not require an immense amount of data is proposed to capture nonlinear structural behavior. In symbolic-based neural networks, symbolic activation functions are capable of finding mathematical expressions to describe the mechanism of input-output relationship, and the hidden state can store the sequence information of nonlinearity. Case studies are carried out for reinforced concrete structures subjected to a series of selected pulse-type ground motions. The results show that the symbolic-based neural network is a promising approach for estimating the nonlinear building responses without a large training dataset.

This paper solves the Sparse Photometric stereo through Lighting Interpolation and Normal Estimation using a generative Network (SPLINE-Net). SPLINE-Net contains a lighting interpolation network to generate dense lighting observations given a sparse set of lights as inputs followed by a normal estimation network to estimate surface normals. Both networks are jointly constrained by the proposed symmetric and asymmetric loss functions to enforce isotropic constrain and perform outlier rejection of global illumination effects. SPLINE-Net is verified to outperform existing methods for photometric stereo of general BRDFs by using only ten images of different lights instead of using nearly one hundred images.