How to scan real scenes

[denghan6688]

How can I scan the 3D model of the environment?

[amir32002]

Below is a brief answer, but please note that in Gibson our goal is turning some already scanned and processed places into a virtual perceptual environment. Our focus is not providing a streamlined tool for scanning new buildings, as that's a sufficiently separate problem. We recommend using the hundreds of spaces already included in Gibson database, as its scale should (hopefully) be sufficient for developing agents that generalize to novel spaces.

However, Gibson's format of the asset data of a space (i.e. each folder in gibson/assets/dataset/) is straightforward. As long as you scan a new space and store it following the same format, Gibson will be able to open it. Some of the requirements are listed below. You can scan the space using one of the existing scanning devices, e.g. Paracosm, DotProduct, Matterport, etc. Gibson is agnostic to scanning hardware.

1. you need to store the RGB images of the space as equirectangular panoramas in panos/rgb folder. Note the resolution and aspect ratio.
2. For each RGB panorama, you need a corresponding depth panorama saved in pano/mist folder.
3. you need a mesh containing the 3D geometry of the space in obj format saved as mesh.obj and a 90 degree rotated version of it saved as mesh_z_up.obj. The first one is used for image rendering and the second in physics engine.
4. you need to save the camera pose information for each image in camera_poses.csv and in individual files in panos/points folder.
5. you need to preserve the naming convention of the existing asset folders.

[denghan6688]

First of all, thank you very much for your answer. I want to know if I still need to retrain the network after scanning the scene myself and if you are open source.

[songWell]

Hi, thank you for your answer. I want to know if there is a depth factor for the corresponding depth panorama ?
#139