.png){kind=icon}
.png){kind=icon}
Return
DKM-Net: self-supervised 3D occupancy prediction via multi-scale dynamic kernel method
.png){kind=icon}
PRE
AI
DOI：10.1007/s00138-026-01806-z.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
To address the challenges of feature sparsity and robustness in object perception for self-supervised 3D occupancy prediction in autonomous driving scenarios, where supervision signals are only depends on multi-camera rendered depth information, existing methods remain constrained by traditional feature extraction networks with limited representation capacity and depth inconsistency during rendering. This paper proposes a multi-scale fusion network composed of Dynamic Kernel Basic modules (DKM-Net). The network enhances backbone feature richness through parallel multi-scale feature processing and feature channel shuffle. Furthermore, we introduce a depth loss function incorporating an embedded Multi-View Consistency Verification (MSV) strategy and kernel density estimation, which leverages multi-frame geometric information to effectively resolve depth inconsistency. By reconstructing the 2D feature extraction architecture and depth estimation constraints, our approach significantly improves both accuracy and efficiency in 3D occupancy prediction. Experimental results demonstrate performance improvements in self-supervised depth estimation and 3D occupancy prediction tasks on the nuScenes dataset.
Keywords:
3D occupancy prediction
Multi scale feature extraction and fusion
Self-supervised depth estimation
Volume rendering
Journal
M
IF:
2.3
Papers:
95
Citations:
2.7K