Efficient Edge-Preserving Multi-View Stereo Network for Depth Estimation
Authors: Wanjuan Su, Wenbing Tao
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that our method achieves state-of-the-art performance with fast inference speed and low memory usage. Notably, our method tops the first place on challenging Tanks and Temples advanced dataset and ETH3D high-res benchmark among all published learning-based methods. |
| Researcher Affiliation | Academia | National Key Laboratory of Science and Technology on Multispectral Information Processing School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, China {suwanjuan, wenbingtao}@hust.edu.cn |
| Pseudocode | No | The paper does not contain any sections or figures explicitly labeled as 'Pseudocode' or 'Algorithm', nor does it present any structured, code-like blocks. |
| Open Source Code | Yes | Code will be available at https://github.com/susuwj/EPNet. |
| Open Datasets | Yes | The DTU dataset (Aanæs et al. 2016), Blended MVS (Yao et al. 2020) dataset, Tanks and Temples dataset (Knapitsch et al. 2017), and ETH3D high-res benchmark (Sch ops et al. 2017) are used. |
| Dataset Splits | Yes | DTU (Aanæs et al. 2016) consists of more than 100 scenes captured under 7 different lighting conditions, which is split into the training set, validation set and evaluation set as (Ji et al. 2017) does and preprocessed as (Yao et al. 2018) does. |
| Hardware Specification | Yes | The experiments are performed on one Ge Force RTX 2080Ti GPU. |
| Software Dependencies | No | The proposed method is implemented by Py Torch (Paszke et al. 2019). |
| Experiment Setup | Yes | During training, the resolution of images for DTU and Blended MVS is set to 640 512, and the number of views N is 5 for DTU and 7 for Blended MVS. The EPNet is composed of 5 stages, the number of depth hypotheses for each stage is set to 32, 16, 8, 8, 8, and the corresponding depth sampling range decays by 0.5 for the second stage and 0.25 for the rest. ... Adam (Kingma and Ba 2014) is used as the optimizer. For the model including the HEPR module, we first train the MSDE alone for 1 epoch, then train the HEPR module alone for 2 epochs to warm up this branch, and finally train the full model for another 9 epochs, and the initial learning rate 0.001 is decreased by half at 8-th, 10-th and 11-th epoch. |