The study introduces AFNet, a novel depth estimation fusion system that adapts to noisy pose settings, outperforming existing methods on KITTI and DDAD datasets. The adaptive fusion module dynamically selects accurate depth between branches based on confidence maps, improving performance under challenging conditions.
The research addresses the limitations of current multi-view systems in autonomous driving scenarios due to noisy poses. By fusing single-view and multi-view depth estimations adaptively, the proposed AFNet achieves state-of-the-art results on challenging benchmarks. The system's robustness is demonstrated through synthetic noise testing and real-world SLAM pose variations.
AFNet integrates single-view features into the multi-view branch, leveraging complementary information for accurate depth estimation. The adaptive fusion module selects reliable depth predictions from both branches based on confidence maps, enhancing accuracy in textureless regions and dynamic object areas. Overall, AFNet demonstrates superior performance in handling noisy poses and dynamic scenes compared to existing methods.
A otro idioma
del contenido fuente
arxiv.org
Ideas clave extraídas de
by JunDa Cheng,... a las arxiv.org 03-13-2024
https://arxiv.org/pdf/2403.07535.pdfConsultas más profundas