The KITTI Vision Benchmark Suite

Feature fusion is the critical component in the
depth completion task. Current approaches mainly
utilize manually designed fusion modules to
construct depth completion networks, but they
generally face the following two problems: 1. The
feature fusion modules at different resolutions
are invariant, requiring the modules to have
multi-scale generalization. 2. The modules
themselves are complex, and additional branches
are needed to enhance features and for post-
processing optimization. Repeated modules and
additional branches lead to network redundancy and
increased computational costs. To address these
challenges, we design a depth completion network
based on neural architecture search. We define the
search space based on cells and employ machine
learning to search for different network
structures at each resolution to construct feature
fusion modules. Meanwhile, we optimize the complex
pruning process in Hierarchical Neural
Architecture Search (HNAS) by defining distinct
cell units to

33.4M