This repository is part of master thesis titled : Synth2Real : 3D-Furniture Reconstruction in Ersatz Environment (S2R:3D-FREE)
This repository is coupled with Deep Learning pipeline which can be seen in /~https://github.com/kartikprabhu20/3dReconstruction
3DScene is a Unity-based pipeline to create synthetic dataset. As a sample the rooms are imported from SceneNet[1] and furnitures are imported from Pix3D[2] The GUI of the application as shown in the figure.
Data Settings: The users can select the catagories/classes of the furnitures seperated by commma(,). The users can also select total images per catagory.
Path Settings: The user has to feed the paths for input which include room and furniture datasets, the path to textures and the destination path.
Texture Directory format:
Texture folder
|_ Furniture1
|_ Furniture2
|_ img1
|_ img2
Room Directory format:
Room folder
|_ roomType1
|_ roomType2
|_ room1.obj
|_ room2.obj
|_room3.obj
Furniture Directory format:
Furniture folder
|_ class1
|_ class2
|_ model_1_folderName
|_ model.obj
|_ model_2_folderName
|_ model.obj
Camera Settings: The parameters include minimum height of the camera, and minimum and maximum distance from the target model.
Light Settings: The lights can be randomized with colors and intensity.
Pipeline Settings: The application supports 4 modes:
- Single room
- Multi-threaded single room
- Multi-objects room
- Manual pipeline
References:
[1] John McCormac, Ankur Handa, Stefan Leutenegger, and Andrew J. Davison. SceneNet RGB-D: Can 5M Synthetic Images Beat Generic ImageNet Pre-training on Indoor Segmentation? In Proceedings of the IEEE International Conference on Computer Vision, 2017.
[2] Xingyuan Sun, Jiajun Wu, Xiuming Zhang, Zhoutong Zhang, Chengkai Zhang, Tianfan Xue, Joshua B. Tenenbaum, and William T. Freeman. Pix3D: Dataset and Methods for Single-Image 3D Shape Modeling. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2018.