search

suryanshkumar / online-joint-depthfusion-and-semantic

A Real-Time Online Learning Framework for Joint 3D Reconstruction and Semantic Segmentation for Indoor Scene.
Other
64 stars 7 forks source link

How to create training Dataset? #3

Open wooca12 opened 2 years ago

wooca12 commented 2 years ago

Hello, I'm interested in your thesis. Currently, only one test data for one room is posted on this Github, and I want to train the model by generating train data. You used Replica and ScanNet datasets as train data, but there was no sample data for ScanNet. I leave a question about how to collect training data.

  1. Can I know the ScanNet data structure used for training? I would appreciate if you could explain the sample data or data conversion code.

  2. I read the Readme file about Replica data in the example data and knew that Replica data should be collected through Habitat-Sim. Before, I had a hard time collecting RGB-D and camera pose as agents in a simulation environment. I'm still exploring how to render Replica method, but I ask you a question to get advice. When collecting consecutive camera poses and RGB-D images, how did you set the agent path? I found .navmesh file in each 3D scenes, did you use this file to collect data? I wonder how you collected Replica data from Hbitat-sim.

Thank you for reading my long question.

cbyan1003 commented 2 years ago

Do you solve your questions? I am also have these questions.