V2XScene: Multi-View Consistent 3D Scene Simulation for Collaborative Perception (WACV2026) скачать в хорошем качестве

V2XScene: Multi-View Consistent 3D Scene Simulation for Collaborative Perception (WACV2026)

This is the official presentation for our WACV 2026 paper: "V2XScene: Multi-View Consistent 3D Scene Simulation for Collaborative Perception." 🚀 Abstract: Data scarcity is a major bottleneck for V2X (Vehicle-to-Everything) collaborative perception. In this work, we introduce V2XScene, a novel 3D scene editing framework designed to generate realistic, label-ready V2X driving scenes. Unlike naive augmentation methods, V2XScene bridges the gap between visual realism and multi-view physical consistency. Key contributions include: Semantic-Guided Generation: Leveraging VQA and Hunyuan3D 2.1 to create high-fidelity 3D vehicle assets. Geometric Consistency: A BEV Layout Optimization Module (BLOOM) that ensures precise alignment between vehicle and infrastructure views. Physical Plausibility: Physics-aware rendering with realistic lighting, shadows, and occlusions. Our experiments show that training with V2XScene-augmented data consistently improves 3D detection performance across state-of-the-art collaborative perception models. 💻 Code: https://github.com/deyang2000/V2XScene.git 🏫 Authors: Yanfei Li, Yi Gong, Yuan Zeng 📍 Conference: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2026)