As a cornerstone for applications such as autonomous driving, 3D urban perception is a burgeoning field of study. Enhancing the performance and robustness of these perception systems is crucial for ensuring the safety of next-generation autonomous vehicles. In this work, we introduce a novel neural scene representation called Street Detection Gaussians (SDGs), which redefines urban 3D perception through an integrated architecture unifying reconstruction and detection. At its core lies the dynamic Gaussian representation, where time-conditioned parameterization enables simultaneous modeling of static environments and dynamic objects through physically constrained Gaussian evolution. The framework’s radar-enhanced perception module learns cross-modal correlations between sparse radar data and dense visual features, resulting in a 22% reduction in occlusion errors compared to vision-only systems. A breakthrough differentiable rendering pipeline back-propagates semantic detection losses throughout the entire 3D reconstruction process, enabling the optimization of both geometric and semantic fidelity. Evaluated on the Waymo Open Dataset and the KITTI Dataset, the system achieves real-time performance (135 Frames Per Second (FPS)), photorealistic quality (Peak Signal-to-Noise Ratio (PSNR) 34.9 dB), and state-of-the-art detection accuracy (78.1% Mean Average Precision (mAP)), demonstrating a $3.8\times$ end-to-end improvement over existing hybrid approaches while enabling seamless integration with autonomous driving stacks.

Image inpainting refers to synthesizing missing content in an image based on known information to restore occluded or damaged regions, which is a typical manifestation of this trend. With the increasing complexity of image in tasks and the growth of data scale, existing deep learning methods still have some limitations. For example, they lack the ability to capture long-range dependencies and their performance in handling multi-scale image structures is suboptimal. To solve this problem, the paper proposes an image inpainting method based on the parallel dual-branch learnable Transformer network. The encoder of the proposed model generator consists of a dual-branch parallel structure with stacked CNN blocks and Transformer blocks, aiming to extract global and local feature information from images. Furthermore, a dual-branch fusion module is adopted to combine the features obtained from both branches. Additionally, a gated full-scale skip connection module is proposed to further enhance the coherence of the inpainting results and alleviate information loss. Finally, experimental results from the three public datasets demonstrate the superior performance of the proposed method.