Name
Abstract | ||
|---|---|---|
Depth completion aims to recover dense depth maps from sparse depth and RGB images. Current methods achieve high accuracy at the cost of large model size and huge computation complexity, which prevent them from wider applications. In this paper, we focus on making two key issues on depth completion - feature extraction and fusion more efficient to achieve superior trade-off in model size and accuracy: (1) we propose efficient dual-branch encoder by exploring data characteristics of different modalities which can greatly reduce the model size and inference time; (2) we propose a dynamic gated fusion module, which is guided by input sparse depth to fuse information of both RGB and sparse depth feature more efficiently by generating dynamic fusing weights. Experiments on KITTI Depth Completion and NYU Depth v2 show that our method achieves 3.5x - 10x speedup against the state-of-art method, 9x param compressing and comparable accuracy compared with state-of-the-art methods, which shows our method achieves good trade-off between performance and speed. |
Year | DOI | Venue |
|---|---|---|
2021 | 10.1007/978-3-030-82153-1_24 | KNOWLEDGE SCIENCE, ENGINEERING AND MANAGEMENT, PT III |
Keywords | DocType | Volume |
Depth completion, Feature fusion, Light-weighted model, Autonomous driving, Robotics | Conference | 12817 |
ISSN | Citations | PageRank |
0302-9743 | 0 | 0.34 |
References | Authors | |
0 | 5 | |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Zhengyang Mu | 1 | 0 | 0.34 |
| Qi Qi | 2 | 210 | 56.01 |
| Jingyu Wang | 3 | 0 | 1.01 |
| Haifeng Sun | 4 | 68 | 27.77 |
| Jianxin Liao | 5 | 457 | 82.08 |