ACDNet: Adaptively Combined Dilated Convolution for Monocular Panorama Depth Estimation

Authors: Chuanqing Zhuang, Zhengda Lu, Yiqun Wang, Jun Xiao, Ying Wang3653-3661

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we conduct depth estimation experiments on three datasets (both virtual and real-world) and the experimental results demonstrate that our proposed ACDNet substantially outperforms the current state-of-the-art (SOTA) methods.
Researcher Affiliation Academia 1 School of Artificial Intelligence, University of Chinese Academy of Sciences 2 College of Computer Science, Chongqing University 3 KAUST zhuangchuanqing19@mails.ucas.ac.cn, {luzhengda,xiaojun,ywang}@ucas.ac.cn, yiqun.wang@cqu.edu.cn
Pseudocode No The paper does not contain a pseudocode block or an algorithm section.
Open Source Code Yes Our codes and model parameters are accessed in https://github.com/zcq15/ACDNet.
Open Datasets Yes We carry out experiments on both virtual and real-world datasets, including Stanford2D3D (Armeni et al. 2017), Matterport3D (Chang et al. 2017), and Structured3D (Zheng et al. 2020).
Dataset Splits Yes For Stanford2D3D and Matterport3D, we follow their official splits with entire panoramic RGB-D pairs to train and test the network. For Structured3D, we just utilize the subset with rawlight illumination and full furniture settings. The subset includes 21, 835 panoramic RGB-D image pairs, and we follow the official scene split for training and testing.
Hardware Specification Yes All experiments were conducted on a server computer equipped with an Intel(R) Xeon(R) Gold 6130 CPU processor, 256GB of RAM, and an NVIDIA TITAN RTX 24GB graphics card.
Software Dependencies No We implement our network on the Py Torch (Paszke et al. 2019) platform. We train our network... with Adam (Kingma and Ba 2015) optimizer respectively. The paper mentions PyTorch and Adam optimizer but does not provide specific version numbers for these software components.
Experiment Setup Yes We train our network for 100 epochs on Stanford2D3D, 60 epochs on Matterport3D, and 60 epochs on Structured3D with Adam (Kingma and Ba 2015) optimizer respectively, the learning rate is set as 1e-4 in all the experiments. Meanwhile, we set the image size as 512 1024 with the batch size of 6 on an NVIDIA TITAN RTX graphics card.