LargeAD: Large-Scale Cross-Sensor Data Pretraining for Autonomous Driving

Autonomous Vehicles (AVs) are advancing rapidly, driven by improvements in intelligent perception and control systems, with a critical focus on reliable object detection in complex environments. Recent research highlights the integration of Vision-Language Models (VLMs) and Large Language Models (LLMs) as pivotal in overcoming existing challenges in multimodal perception and contextual reasoning.

LATTICE has introduced a new framework for high-fidelity 3D asset generation, addressing the challenges of predicting spatial structures and geometric surfaces in 3D models. This framework utilizes VoxSet, a semi-structured representation that compresses 3D assets into latent vectors, enhancing efficiency and scalability in 3D generation compared to traditional 2D methods.

A comprehensive survey titled '3D and 4D World Modeling' has been published, addressing the critical role of world modeling in AI research. It highlights the need for standardized definitions and taxonomies in the field, focusing on 3D and 4D representations such as RGB-D imagery, occupancy grids, and LiDAR point clouds, which have been underrepresented in previous studies.

ShelfGaussian has been introduced as an open-vocabulary multi-modal Gaussian-based framework for 3D scene understanding, leveraging off-the-shelf vision foundation models to enhance performance and efficiency in various scene understanding tasks. This framework addresses limitations of existing methods by enabling Gaussians to query features from multiple sensor modalities and optimizing them at both 2D and 3D levels.

GaussianBlender has been introduced as a groundbreaking framework for text-driven 3D stylization, enabling instant edits at inference by utilizing structured, disentangled latent spaces derived from spatially-grouped 3D Gaussians. This innovation addresses the inefficiencies of traditional text-to-3D methods that require extensive optimization and often result in multi-view inconsistencies.

GT23D-Bench has been introduced as a comprehensive benchmark for General Text-to-3D (GT23D) generation, focusing on synthesizing 3D content from textual descriptions without the need for model re-optimization. This shift aims to enhance efficiency and generalization in 3D content creation, addressing the limitations of existing per-scene approaches.

A systematic investigation has been conducted to evaluate how different LiDAR-to-image projections impact metric place recognition when integrated with advanced vision foundation models. The study introduces a modular retrieval pipeline that isolates the effects of 2-D projections, identifying key characteristics that enhance discriminative power and robustness in various environments.

A new framework named BEVDilation has been introduced, focusing on the integration of LiDAR and camera data for enhanced 3D object detection. This approach emphasizes LiDAR information to mitigate performance degradation caused by the geometric discrepancies between the two sensors, utilizing image features as implicit guidance to improve spatial alignment and address point cloud limitations.