This is the dataset release of our paper Through the Looking Glass: Neural 3D Reconstruction of Transparent Shapes, CVPR 2020. Please consider citing this paper if you find this dataset useful in your project. Please contact us if you have any questions or issues.
Followings are the links to our dataset, to understand the data structure of our dataset, please refer to the dataset creation repository from this link and the network repository from this link, which will tell you how the data are created and how to load the data for training. We also provide our final reconstruction results of 10-view, 20-view and 5-view reconstruction so that future researchers can compare with our method easily.
- Shapes
- The geometry, camera position and scene configuration files used to create the dataset.
- Images5
- The rendered images, two-bounce normal and the final reconstructed meshes of the 5-view reconstruction.
- Images10
- The rendered images, two-bounce normal and the final reconstructed meshes of the 10-view reconstruction.
- Images20
- The rendered images, two-bounce normal and the final reconstructed meshes of the 20-view reconstruction.
- Envmap
- We use the Laval Indoor Lighting Dataset to render our data. Unfortunately, we are not supposed to redistribute the dataset. We have actually rescaled and renamed the environment maps when creating our data. The
infoList.datandmapEnvmaps.pycan be used to map the original environment maps to our environment maps used for rendering.
- We use the Laval Indoor Lighting Dataset to render our data. Unfortunately, we are not supposed to redistribute the dataset. We have actually rescaled and renamed the environment maps when creating our data. The
To reproduce the quantitative number of shape reconstruction in our paper, you can download the dataset and run testMesh.py from this link with --dataRoot and --camNum set properly. Please let us know if you have any questions running the code.