The generation of temporally consistent, high-fidelity driving videos over extended horizons presents a fundamental challenge in autonomous driving world modeling. Existing approaches often suffer from error accumulation and feature misalignment due to inadequate decoupling of spatio-temporal dynamics and limited cross-frame feature propagation mechanisms. To address these limitations, we present STAGE (Streaming Temporal Attention Generative Engine), a novel auto-regressive framework that pioneers hierarchical feature coordination and multi-phase optimization for sustainable video synthesis. To achieve high-quality long-horizon driving video generation, we introduce Hierarchical Temporal Feature Transfer (HTFT) and a novel multi-stage training strategy. HTFT enhances temporal consistency between video frames throughout the video generation process by modeling the temporal and denoising process separately and transferring denoising features between frames. The multi-stage training strategy is to divide the training into three stages, through model decoupling and auto-regressive inference process simulation, thereby accelerating model convergence and reducing error accumulation. Experiments on the Nuscenes dataset show that STAGE has significantly surpassed existing methods in the long-horizon driving video generation task. In addition, we also explored STAGE’s ability to generate unlimited-length driving videos. We generated 600 frames of high-quality driving videos on the Nuscenes dataset, which far exceeds the maximum length achievable by existing methods. Our project homepage:https://4dvlab.github.io/STAGE/

In recent years, vision-centric Bird's Eye View (BEV) perception has garnered significant interest from both industry and academia due to its inherent advantages, such as providing an intuitive representation of the world and being conducive to data fusion. The rapid advancements in deep learning have led to the proposal of numerous methods for addressing vision-centric BEV perception challenges. However, there has been no recent survey encompassing this novel and burgeoning research field. To catalyze future research, this paper presents a comprehensive survey of the latest developments in vision-centric BEV perception and its extensions. It compiles and organizes up-to-date knowledge, offering a systematic review and summary of prevalent algorithms. Additionally, the paper provides in-depth analyses and comparative results on various BEV perception tasks, facilitating the evaluation of future works and sparking new research directions. Furthermore, the paper discusses and shares valuable empirical implementation details to aid in the advancement of related algorithms.

Abstract This study proposes a self‐supervised method to reconstruct 3D limbic structures from 2D landmarks extracted from a single view. The loss of self‐consistency can be reduced by performing a random orthogonal projection of the reconstructed 3D structure. Thus, the training process can be self‐supervised by using geometric self‐consistency in the reconstruction–projection–reconstruction process. The self‐supervised network mainly consists of graph convolution and Transformer encoders. This network is called the SS‐Graphformer. By adding a discriminator, the SS‐Graphformer is used as a generator to form a Wasserstein Generative Adversarial Network architecture with a Gradient Penalty to improve the accuracy of the reconstruction. It is experimentally demonstrated that the addition of the 2D structure discriminator can significantly improve the accuracy of the reconstruction.

As one of the most fascinating features of ultra-high definition television (UHDTV), high dynamic range (HDR) can significantly improve visual experience compared with standard dynamic range (SDR). During the transition from SDR to HDR, it is possible to mix SDR and HDR cameras or contents in the process of production. Consequently, various conversions, including SDR-HDR conversion, HDR-SDR conversion, SDR-HDR-SDR round-trip conversion and etc., are necessary to combine the contents of different formats in the same workflow, in which it is challenging to keep consistent video quality after complex conversions. This paper analyzes scene-referred and display-referred methods of conversions between SDR and HDR in the practice of UHDTV production and conducts experiments to evaluate the conversion performance. Based on the experimental results, suggestions on conversions between SDR and HDR are provided for practical application scenarios.

Compared to earlier multistage frameworks using CNN features, recent end-to-end deep approaches for fine-grained recognition essentially enhance the mid-level learning capability of CNNs. Previous approaches achieve this by introducing an auxiliary network to infuse localization information into the main classification network, or a sophisticated feature encoding method to capture higher order feature statistics. We show that mid-level representation learning can be enhanced within the CNN framework, by learning a bank of convolutional filters that capture class-specific discriminative patches without extra part or bounding box annotations. Such a filter bank is well structured, properly initialized and discriminatively learned through a novel asymmetric multi-stream architecture with convolutional filter supervision and a non-random layer initialization. Experimental results show that our approach achieves state-of-the-art on three publicly available fine-grained recognition datasets (CUB-200-2011, Stanford Cars and FGVC-Aircraft). Ablation studies and visualizations are provided to understand our approach.

Deep Convolutional Neural Networks (CNN) enforce supervised information only at the output layer, and hidden layers are trained by back propagating the prediction error from the output layer without explicit supervision. We propose a supervised feature learning approach, Label Consistent Neural Network, which enforces direct supervision in late hidden layers in a novel way. We associate each neuron in a hidden layer with a particular class label and encourage it to be activated for input signals from the same class. More specifically, we introduce a label consistency regularization called "discriminative representation error" loss for late hidden layers and combine it with classification error loss to build our overall objective function. This label consistency constraint alleviates the common problem of gradient vanishing and tends to faster convergence, it also makes the features derived from late hidden layers discriminative enough for classification even using a simple k-NN classifier. Experimental results demonstrate that our approach achieves state-of-the-art performances on several public datasets for action and object category recognition.

Fine-grained classification involves distinguishing between similar sub-categories based on subtle differences in highly localized regions, therefore, accurate localization of discriminative regions remains a major challenge. We describe a patch-based framework to address this problem. We introduce triplets of patches with geometric constraints to improve the accuracy of patch localization, and automatically mine discriminative geometrically-constrained triplets for classification. The resulting approach only requires object bounding boxes. Its effectiveness is demonstrated using four publicly available fine-grained datasets, on which it outperforms or achieves comparable performance to the state-of-the-art in classification.

Latent Low-Rank Representation (Lat LRR) has the empirical capability of identifying "salient" features. However, the reason behind this feature extraction effect is still not understood. Its optimization leads to non-unique solutions and has high computational complexity, limiting its potential in practice. We show that Lat LRR learns a transformation matrix which suppresses the most significant principal components corresponding to the largest singular values while preserving the details captured by the components with relatively smaller singular values. Based on this, we propose a novel feature extraction method which directly designs the transformation matrix and has similar behavior to Lat LRR. Our method has a simple analytical solution and can achieve better performance with little computational cost. The effectiveness and efficiency of our method are validated on two face recognition datasets.

This paper considers a basic MIMO information-energy broadcast system, where a multi-antenna transmitter transmits information and energy simultaneously to a multi-antenna information receiver and a dual-functional multi-antenna energy receiver which is also capable of decoding information. Due to the open nature of wireless medium and the dual purpose of information and energy transmission, secure information transmission while ensuring efficient energy harvesting is a critical issue for such a broadcast system. Providing that physical layer security techniques are adopted for secure transmission, we study beamforming design to maximize the achievable secrecy rate subject to a total power constraint and an energy harvesting constraint. First, based on semidefinite relaxation, we propose global optimal solutions to the secrecy rate maximization (SRM) problem in the single-stream case and a specific full-stream case. Then, we propose inexact block coordinate descent (IBCD) algorithm to tackle the SRM problem of general case with arbitrary number of streams. We prove that the IBCD algorithm can monotonically converge to a Karush-Kuhn-Tucker (KKT) solution to the SRM problem. Furthermore, we extend the IBCD algorithm to the joint beamforming and artificial noise design problem. Finally, simulations are performed to validate the effectiveness of the proposed beamforming algorithms.

The transmission rate, delivery reliability, and network lifetime are three fundamental but conflicting design objectives in energy-constrained wireless sensor networks (WSNs). In this paper, based on stochastic network utility maximization framework, we address the optimal rate-reliability-lifetime tradeoff with time-varying channel capacity constraint, reliability constraint, and energy constraint. By introducing the weight parameters, we combine the optimization objectives of rate, reliability, and lifetime into a single objective to characterize the tradeoff among them. However, the optimization formulation of the rate-reliability-reliability tradeoff is neither separable nor convex. Through a series of transformations, a separable problem is derived, and an efficient distributed stochastic subgradient algorithm is proposed via dual decomposition and stochastic subgradient techniques. It is proved that the proposed algorithm can converge to the global optimum with probability one. Numerical examples confirm its convergence. In addition, numerical examples investigate the impact of weight parameters on the rate utility, reliability utility, and network lifetime, which provide guidance to properly set the value of weight parameters for a desired performance of WSNs according to the realistic application's requirements.

The Modulated Wideband Converter (MWC) is a recently proposed analog-to-digital converter (ADC) based on Compressive Sensing (CS) theory. Unlike conventional ADCs, its quantization reference voltage, which is important to the system performance, does not equal the maximum amplitude of original analog signal. In this paper, the quantization reference voltage of the MWC is theoretically analyzed and the conclusion demonstrates that the reference voltage is proportional to the square root of q, which is a trade-off parameter between sampling rate and number of channels. Further discussions and simulation results show that the reference voltage is proportional to the square root of Nq when the signal consists of N narrowband signals.

Using cyclic prefix (CP), the transmission schemes, orthogonal frequency-division multiplexing (OFDM), single carrier block transmission, and time reversal, can be unified as linear block precoding. Considering frequency-selective channels, this paper studies linear block precoding for CP-based multi-point transmission with per-antenna power constraints (PAPCs) under capacity maximization and mean-square-error (MSE) minimization criteria. We show that, the optimal precoders for both criteria could be, but not necessarily, in the form of OFDM transmission (i.e., an inverse discrete fourier transformation (IDFT) matrix multiplying a complex diagonal matrix). Based on the optimal precoder structure, the two problems are simplified to two matrix-free optimization problems for which we prove strong duality holds. Moreover, it is shown that the dual problems can be equivalent to two unconstrained convex optimization problems. Efficient optimum precoding algorithms are proposed for both problems. Simulation results show that the maximum capacity (or minimum MSE) in the PAPC case almost coincides with that in the sum power constraint (SPC) case when the power budgets for each antenna are equal, but a capacity (or MSE) gap exists between the two power constraint cases when the power budgets for each antenna are different.

Ad Hoc networks are characterized as fast time-varying. Thus, fast distributed algorithm to implement self-management is indispensable, especially for QoS support. In this paper, we propose rate control with QoS support in Ad Hoc networks based on primal-dual interior-point method. We apply Gaussian belief propagation algorithm to compute the Newton step. For implementing distributed computation in practical network, we design the mapping rules between GaBP-mapping network and the practical networks. Finally, the simulation results show that the proposed algorithm has favorable performance, including fast convergence, robustness and scalability.