wacv2026

Overview

Hybrid rendering method combining neural reconstruction with physics-based rendering for autonomous driving development. Integrated into aiSim simulator, supports multiple sensor modalities, camera models, and predicts segmentation, normals, depth. Validated via downstream tasks, HiL, and closed-loop tests.

Key features

• ▸Hybrid rendering combining neural reconstruction with physics-based rendering
• ▸Supports multiple sensor modalities (LiDAR, radar target lists)
• ▸Supports different camera models (e.g., fisheye)
• ▸Accounts for camera exposure mismatches
• ▸Predicts segmentation masks, surface normals, and depth maps
• ▸Validated through downstream tasks, HiL experiments, and closed-loop ADAS/AD tests
• ▸NeRF2GS method enables high-quality novel view synthesis after extreme viewpoint changes
• ▸Works in different operational design domains (urban, highway, proving ground)
• ▸Can be applied to public datasets like Waymo
• ▸Supports equirectangular camera model with 3DGS