This repo contains the modified version of Replica Dataset to generate the training and testing dataset used in MatryODShka paper. This is a companion to the training code repo.
The original ReplicaRenderer and ReplicaViewer remain the same. See original repo for basic usage. See below for 360 dataset generation.
- Git Clone this repo.
- Run
git submodule update --init --recursiveto install 3rd party packages. - Run
./download.shto download the mesh files. - Run
./build.shto build all executables.
To generate the training and testing dataset, run
./build/ReplicaSDK/ReplicaRendererDataset dataset/scene_name/mesh.ply
dataset/scene_name/textures dataset/scene_name/glass.sur camera_parameters[file.txt / n]
spherical[y/n] output/dir/ output_width output_height pro2[globs/pro2pos.txt / n]
The data generation takes in a text file specifying the camera position, ods baseline and target camera positions for each navigable position within the scene. A single line in the input text file (camera_parameters.txt) is formatted as:
camera_position_x camera_position_y camera_position_z ods_baseline
target1_offset_x target1_offset_y target1_offset_z
target2_offset_x target2_offset_y target2_offset_z
target3_offset_x target3_offset_y target3_offset_z
The existing text files contain navigable positions within each scene, sampled with Habitat-SIM. Find all the existing text files in glob/.
This repo also supports video dataset rendering specifically.
./build/ReplicaSDK/ReplicaVideoRenderer path/to/scene/mesh.ply
path/to/scene/textures path/to/scene/glass.sur camera_parameters.txt
spherical[y/n] output/dir/ width height
The difference is the format of a single line in input text file for video rendering is:
camera_position_x camera_position_y camera_position_z
lookat_x lookat_y lookat_z ods_baseline
rotation_x rotation_y rotation_z target_x target_y target_z
To generate a video path, one can use glob/gen_video_path.py. See glob/example_script for an example of straight path generation.