As immersive experiences become increasingly popular, panoramic video has garnered significant attention in both research and applications. The high cost associated with capturing panoramic video underscores the need for efficient prompt-based generation methods. Although recent text-to-video (T2V) diffusion techniques have shown potential in standard video generation, they face challenges when applied to panoramic videos due to substantial differences in content and motion patterns. In this paper, we propose PanoDiT, a framework that utilizes the Diffusion Transformer (DiT) architecture to generate panoramic videos from text descriptions. Unlike traditional methods that rely on UNet-based denoising, our method leverages a transformer architecture for denoising, incorporating both temporal and global attention mechanisms. This ensures coherent frame generation and smooth motion transitions, offering distinct advantages in long-horizon generation tasks. To further enhance motion and consistency in the generated videos, we introduce DTM-LoRA and two panoramic-specific losses. Compared to previous methods, our PanoDiT achieves state-of-the-art performance across various evaluation metrics and user study, with code is available in the supplementary material.

Multimodal Large Language Models (MLLMs) have achieved notable success in visual instruction tuning, yet their inference is time-consuming due to the auto-regressive decoding of Large Language Model (LLM) backbone. Traditional methods for accelerating inference, including model compression and migration from language model acceleration, often compromise output quality or face challenges in effectively integrating multimodal features. To address these issues, we propose AASD, a novel framework for Accelerating inference with refined KV Cache and Aligning speculative decoding in MLLMs. Our approach leverages the target model’s cached KeyValue (KV) pairs to extract vital information for generating draft tokens, enabling efficient speculative decoding. To reduce the computational burden associated with long multimodal token sequences, we introduce a KV Projector to compress the KV Cache while maintaining representational fidelity. Additionally, we design a Target-Draft Attention mechanism that optimizes the alignment between the draft model and the target model, achieving the benefits of real inference scenarios with minimal computational overhead. Extensive experiments on mainstream MLLMs demonstrate that our method achieves up to a $2 \times$ inference speedup without sacrificing accuracy. This study not only provides an effective and lightweight solution for accelerating MLLM inference but also introduces a novel alignment strategy for speculative decoding in multimodal contexts, laying a strong foundation for future research in efficient MLLMs. Code is availiable at https://github.com/transcend-0/ASD

Forward collision warning (FCW) systems are available in many new vehicles and are becoming increasingly popular. It is important to clearly understand their effectiveness in reducing collision risks, as well as their potential negative impacts on driving safety and volatility. This research aims to answer this question using real-world connected vehicles (CV) data collected under naturalistic settings. We extract 2,332 FCW events from the New York City CV Pilot Deployment dataset, including 1,326 events in the treatment group (warning issued) and 1,006 in the control group (no warning issued). From these FCW events, eleven volatility and nine safety variables are proposed and calculated, considering the interactions between the following and leading vehicles. These variables are further filtered using an ensemble variable elimination and selection method based on Variance Inflation Factors, Person correlation, and Akaike Information Criterion. Three binary logit models are developed for modeling volatility, safety, and both volatility and safety, respectively. These models are compared based on their log-likelihood values and the margin effects (MEs) of variables. The results show that FCW systems significantly enhance driving safety by reducing the time vehicles spend in high-risk conditions (ME: −4.59%) and the period of harsh braking to avoid collisions (ME: −5.27%) through issuing audible warnings. Despite increases in driving volatility, the overall benefits of FCW on safety outweigh the risks, with a benefit-to-risk ratio of 48.72%, resulting in a favorable net effect on driving dynamics from a statistical perspective.

Forward-looking multi-input multi-output synthetic aperture radar (FL-MIMO-SAR) is the state-of-the-art technique to achieve the high-resolution image of the front area. However, the performance of FL-MIMO-SAR is primary limited by left-right Doppler ambiguity resolution. To address this issue, a novel space-time cascade imaging framework is proposed. Concretely, the Doppler-division multiplexing MIMO (DDM-MIMO) scheme is adopted to achieve the waveform orthogonality and the back-projection (BP) time-domain algorithm is applied to obtain high-resolution SAR images. Furthermore, the effect of array errors is analyzed and calibrated, which is essential for subsequent Doppler ambiguity resolution using the linearly constrained minimum variance (LCMV) beamforming. Finally, experimental analysis is performed to confirm the effectiveness of the proposal.

Numerous studies suggest that the visual system uses both phase- and position-shift receptive field (RF) mechanisms for the processing of binocular disparity. Although the difference between these two mechanisms has been analyzed before, previous work mainly focused on disparity tuning curves instead of population responses. However, tuning curve and population response can exhibit different characteristics, and it is the latter that determines disparity estimation. Here we demonstrate, in the framework of the disparity energy model, that for relatively small disparities, the population response generated by the phase-shift mechanism is more reliable than that generated by the position-shift mechanism. This is true over a wide range of parameters, including the RF orientation. Since the phase model has its own drawbacks of underestimating large stimulus disparity and covering only a restricted range of disparity at a given scale, we propose a coarse-to-fine algorithm for disparity computation with a hybrid of phase-shift and position-shift components. In this algorithm, disparity at each scale is always estimated by the phase-shift mechanism to take advantage of its higher reliability. Since the phase-based estimation is most accurate at the smallest scale when the disparity is correspondingly small, the algorithm iteratively reduces the input disparity from coarse to fine scales by introducing a constant position-shift component to all cells for a given location in order to offset the stimulus disparity at that location. The model also incorporates orientation pooling and spatial pooling to further enhance reliability. We have tested the algorithm on both synthetic and natural stereo images and found that it often performs better than a simple scale-averaging procedure.