Meaning of parameters in info.json

[watashihame]

Thank you for releasing the code of your extraordinary work.

I am trying to reconstruct smoke from my data with your code. But I am stuck on creating the info.json of my data.

Such as parameters in data/Sphere/info.json

    "voxel_scale": 12.0064,
    "voxel_matrix": [
        [
            0.4000000059604645,
            0.0,
            0.0,
            -2.419999837875366
        ],
        [
            0.0,
            -1.7484556025237907e-08,
            -0.4000000059604645,
            2.460000514984131
        ],
        [
            0.0,
            0.4000000059604645,
            -1.7484556025237907e-08,
            -2.7099997997283936
        ],
        [
            0.0,
            0.0,
            0.0,
            1.0
        ]
    ],
    "render_center":[
        0.0, 
        0.0, 
        0.0
    ],
    "near":2.0,
    "far":6.0,
    "phi":-30.0,
    "rot":"Z"

Could you tell me what these parameters refer to? Thank you.

[RachelCmy]

Hi,

parameters in data/Sphere/info.json are set according to our synthetic data based on Blender and Openvdb. "voxel_scale" is the size of the voxel in the world coord. For example, our synthetic Sphere data use:

• voxel grid in the resolution of 256^3
• voxel grid with cell length 0.0469 ( which was set with vscale= 0.0469 instead of: https://github.com/RachelCmy/pinf_smoke/blob/4b7cb02c2890ab42124aab9460eddbe108bc4666/data/Game/scripts/manta2vdb.py#L7 ) So its voxel_scale is 256*0.0469= 12.0064 on x,y,and z.
• The voxel_scale can vary on x,y, and z. (for e.g., the Game scene has "voxel_scale" = [ 24.0128, 20.2608, 19.1352])

"voxel_matrix" is the transformation matrix of the smoke data in blender, which is set by bpy.context.scene.objects['our_fluid_data'].matrix_world

"voxel_scale" and "voxel_matrix" are used to transfer coord. between voxel space and the world space. There is little explanation at: https://github.com/RachelCmy/pinf_smoke/blob/4b7cb02c2890ab42124aab9460eddbe108bc4666/run_pinf_helpers.py#L487

[RachelCmy]

"render_center", "near", "far", "phi","rot" are used to set the poses for spherical rendering (https://github.com/RachelCmy/pinf_smoke/blob/e8999f902f31ef33ff8bbc6f07bfde7e216f6fce/load_pinf.py#L160)

[watashihame]

Thank you for the answer.

But how should I set these values for a series of real captured data when I don't know the voxel_scale and voxel_transform? Like ones in data/ScalarReal/info.json https://github.com/RachelCmy/pinf_smoke/blob/4b7cb02c2890ab42124aab9460eddbe108bc4666/data/ScalarReal/info.json#L189-L219

Or just set them into some default values?

[RachelCmy]

Hi,

for real data, voxel_transform can be set to an identity matrix. render_center should be the center of the volume ( some positions near the intersection point of some camera directions). voxel_scale determines the size of the visible volumetric area (maybe try something between distance_from_camera_to_render_center * tan(.5 * camera_angle_x) and 2*distance_from_camera_to_render_center * tan(.5 * camera_angle_x))

Networks are only trained by sampling inside the volumetric area. So if it is too little, the networks are not able to reconstruct things outside. If it is too large, the reconstruction will be blurry since the sampling is too sparse.