Iterative Geometry-Aware Cross Guidance Network for Stereo Image Inpainting
Authors: Ang Li, Shanshan Zhao, Zhang Qingjie, Qiuhong Ke
IJCAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that our proposed network outperforms the latest stereo image inpainting model and state-of-the-art single image inpainting models. Experiments on real data demonstrate that our proposed IGGNet outperforms state-of-the-art single image inpainting methods and stereo image inpainting methods. |
| Researcher Affiliation | Collaboration | Ang Li1 , Shanshan Zhao2 , Zhang Qingjie1 and Qiuhong Ke3 1Aviation University of Air Force 2JD Explore Academy, JD.COM 3The University of Melbourne |
| Pseudocode | Yes | Algorithm 1 illustrates the procedure of the ICG strategy. |
| Open Source Code | No | The paper mentions using 'released code' for baseline models but does not provide a specific link or explicit statement about the availability of their own source code for the proposed IGGNet. |
| Open Datasets | Yes | We train and evaluate our model on two wellknown stereo image datasets: KITTI2015 [Mayer et al., 2016] and MPI-Sintel [Butler et al., 2012]. |
| Dataset Splits | Yes | For the KITTI2015 dataset, we use the original training and testing splits. For the MPI-Sintel dataset, we select 16 video directories for training, and the other 7 for testing. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. It only mentions 'Details of network structures are provided in the supplementary material' which is not about hardware. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers (e.g., Python version, library versions, specific frameworks). |
| Experiment Setup | Yes | All training images are randomly cropped to 256 256 for every epoch. Each testing image in the KITTI2015 dataset is horizontally split to 4 images and resize them to 256 256, while that in the MPI-Sintel dataset is directly resized to 256 256. In our experiments, we define the maximum image disparity to be 192. Since we apply GAA module on the 1/4 scale level, D is set as 48. We use three different mask ratios (mask area relative to the entire image): 0 20%, 20 40% and 40 60%. |