HR-Depth: High Resolution Self-Supervised Monocular Depth Estimation
Authors: Xiaoyang Lyu, Liang Liu, Mengmeng Wang, Xin Kong, Lina Liu, Yong Liu, Xinxin Chen, Yi Yuan2294-2301
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments In this section, we validate that (1) our redesigned skip-connection can improve the results, especially predicting sharper edges, (2) the f SE module can significantly reduce parameters and improve accuracy, and (3) the design method we propose can easily obtain high precision lightweight network. We evaluate our models on the KITTI dataset(Geiger et al. 2012), to allow comparison with previous published monocular methods. |
| Researcher Affiliation | Collaboration | Xiaoyang Lyu 1, Liang Liu 1, Mengmeng Wang 1, Xin Kong 1, Lina Liu 1, Yong Liu* 1, Xinxin Chen 1, Yi Yuan 2 1 Zhejiang University 2 Fuxi AI Lab, Net Ease {shawlyu, leonliuz, mengmengwang, xinkong, linaliu}@zju.edu.cn, yongliu@iipc.zju.edu.cn, cxx-compiler@zju.edu.cn, yuanyi@corp.netease.com |
| Pseudocode | No | The paper describes the architecture and components of HR-Depth but does not include pseudocode or a clearly labeled algorithm block. |
| Open Source Code | Yes | All codes and models will be available at https: //github.com/shaw Lyu/HR-Depth. |
| Open Datasets | Yes | We evaluate our models on the KITTI dataset(Geiger et al. 2012), to allow comparison with previous published monocular methods. The KITTI benchmark (Geiger et al. 2012) is most widely used for depth evaluation. |
| Dataset Splits | Yes | We adopt the data split of (Eigen et al. 2015), and removed the static frames followed by (Zhou et al. 2017). Ultimately, we used 39810 images for training, 4424 for validation and 697 for evaluation. |
| Hardware Specification | Yes | We implement our models on Py Torch(Paszke et al. 2017) and train them on one Telsa V100 GPU. |
| Software Dependencies | No | The paper mentions 'Py Torch' but does not specify its version number or any other software dependencies with their versions. |
| Experiment Setup | Yes | We use the Adam Optimizer(Kingma et al. 2014) with β1 = 0.9, β2 = 0.999. The Depth Net and Pose Net are trained for 20 epochs, with a batch size of 12. The initial learning rates for both network are 1 10 4 and decayed after 15 epochs by factor of 10. The training sequences are consist of three consecutive images. We set the SSIM weight to α = 0.85 and smooth loss weight to λ = 1 10 3. |