Create Multi-View Stereo sequence

[rossimattia]

• I would like to generate a multi-view stereo sequence, i.e., images of the same scenes captured from different points of view. Is this possible using the gin configurations?

• As a workaround for the multi-view stereo dataset generation, I thought about generating a monocular video with a command similar to the following one: python -m infinigen.datagen.manage_jobs --output_folder outputs/my_videos --num_scenes 1 \ --pipeline_config monocular_video cuda_terrain opengl_gt --cleanup big_files --config video high_quality_terrain However, the default frame rate of 24fps is a huge waste of resources, as I would need a very sparse video (e.g., one frame every few seconds, such that the camera can move significantly between once frame and the next one), but in the documentation it is not clear whether it is possible to change the frame rate.

• Since each scene is generate at random, how is the camera trajectory decided in the video? Is it possible to define the camera trajectory after the scene has been generated?

Any suggestion would be very welcome. Thank you.

[araistrick]

Hello, you can change the frame rate with the commandline override -p execute_tasks.fps=1, and you can similarly change the camera speed to make the frames as sparse as you like with AnimPolicyRandomWalkLookaround.speed = ("uniform", 3, 10) (this is the default speed - you may want to increase it)

RE camera trajectories, yes they can be decided however you like after the scene is generated. We provide code to randomly generate them called here, you are free to modify this code or write your own entirely custom script.