In this repository, we only support matterport3D dataset for training and inference.
Here we'll introduce the data orginization for further experiments.
2022-09-26: All the required data in this repo can be downloaded once you sign this Google Form. Note that the released dataset is the resized version. You can see this issue to see more information if needed.
dataset/
├── 17DRP5sb8fy
│ ├── estimate_boundary
│ │ ├── resize_00ebbf3782c64d74aaf7dd39cd561175_d1_0_bound_est.png
│ │ └── ...
│ ├── estimate_normal
│ │ ├── resize_00ebbf3782c64d74aaf7dd39cd561175_d1_0_normal_est.png
│ │ └── ...
│ ├── render_depth
│ │ ├── resize_00ebbf3782c64d74aaf7dd39cd561175_d0_0_mesh_depth.png
│ │ └── ...
│ ├── sobel_edge
│ │ ├── resize_00ebbf3782c64d74aaf7dd39cd561175_d0_0_mesh_depth.png
│ │ └── ...
│ ├── undistorted_color_images
│ │ ├── resize_00ebbf3782c64d74aaf7dd39cd561175_i0_0.jpg
│ │ └── ...
│ └── undistorted_depth_images
│ ├── resize_00ebbf3782c64d74aaf7dd39cd561175_d0_0.png
│ └── ...
└─── 1LXtFkjw3qL
└── ...Following matterport3D dataset, the data orginization in this repository is shown above. In our experiment, we resize all images to 256x320 for convenience. Let's introduce each directory respectively.
- undistorted_color_images : resized original rgb images
- undistorted_depth_images : resized original depth images
- render_depth : ground truth depth map generated by multiview-reconstruction method (also in yindaZ's work)
- estimate_boundary : occlusion boundary pre-generated by yindaZ's model
- estimate_normal : surface normal pre-generated by yindaZ's model
- sobel_edge : render_depth's boundary