Skip to content

Latest commit

 

History

History
135 lines (93 loc) · 9.85 KB

HelloRoom.md

File metadata and controls

135 lines (93 loc) · 9.85 KB
Raw

Hello Room: Generate your first Infinigen-Indoors scene

Generate a scene step-by-step

Infinigen has distinct scene generation & rendering stages. We typically run these automatically for you (skip to Generate scenes automatically

Generate a blender file

First, run ONE command of your choosing from the block below. This will generate a 3D blender file for use in the subsequent steps.

NOTE: fast_solve.gin runs the system for fewer solving steps, which sacrifices quality for speed. Remove t his to get a more complex & realistic arrangement. You can also remove compose_indoors.terrain_enabled=False to add a realistic terrain background (provided you installed terrain)

# Diningroom, single room only, first person view (~8min CPU runtime)
python -m infinigen_examples.generate_indoors --seed 0 --task coarse --output_folder outputs/indoors/coarse -g fast_solve.gin singleroom.gin -p compose_indoors.terrain_enabled=False restrict_solving.restrict_parent_rooms=\[\"DiningRoom\"\]

# Bathroom, single room only, first person view (~13min CPU runtime)
python -m infinigen_examples.generate_indoors --seed 0 --task coarse --output_folder outputs/indoors/coarse -g fast_solve.gin singleroom.gin -p compose_indoors.terrain_enabled=False restrict_solving.restrict_parent_rooms=\[\"Bathroom\"\]

# Bedroom, single room only, first person view (~10min CPU runtime)
python -m infinigen_examples.generate_indoors --seed 0 --task coarse --output_folder outputs/indoors/coarse -g fast_solve.gin singleroom.gin -p compose_indoors.terrain_enabled=False restrict_solving.restrict_parent_rooms=\[\"Bedroom\"\]

# Kitchen, single room only, first person view (~10min runtime, CPU only)
python -m infinigen_examples.generate_indoors --seed 0 --task coarse --output_folder outputs/indoors/coarse -g fast_solve.gin singleroom.gin -p compose_indoors.terrain_enabled=False restrict_solving.restrict_parent_rooms=\[\"Kitchen\"\]

# LivingRoom, single room only, first person view (~11min runtime, CPU only)
python -m infinigen_examples.generate_indoors --seed 0 --task coarse --output_folder outputs/indoors/coarse -g fast_solve.gin singleroom.gin -p compose_indoors.terrain_enabled=False restrict_solving.restrict_parent_rooms=\[\"LivingRoom\"\]

# Floor layout, overhead view, no objects (~34 second runtime, CPU only):
python -m infinigen_examples.generate_indoors --seed 0 --task coarse --output_folder outputs/indoors/coarse -g no_objects.gin overhead.gin -p compose_indoors.terrain_enabled=False

# Single random room with objects, overhead view (~11min. runtime CPU only):
python -m infinigen_examples.generate_indoors --seed 0 --task coarse --output_folder outputs/indoors/coarse -g fast_solve.gin overhead.gin singleroom.gin -p compose_indoors.terrain_enabled=False compose_indoors.overhead_cam_enabled=True restrict_solving.solve_max_rooms=1 compose_indoors.invisible_room_ceilings_enabled=True compose_indoors.restrict_single_supported_roomtype=True

# Whole apartment with objects, overhead view:
python -m infinigen_examples.generate_indoors --seed 0 --task coarse --output_folder outputs/indoors/coarse -g fast_solve.gin overhead.gin studio.gin -p compose_indoors.terrain_enabled=False

Once complete, you can inspect / fly around outputs/indoors/coarse/scene.blend in the blender UI:

python -m infinigen.launch_blender outputs/indoors/coarse/scene.blend

You may be prompted to revisit our Installation.md if blender is not yet installed.

Render image and ground truth

Next, run the commands below to render an RGB image and ground truth:

# Render RGB images
python -m infinigen_examples.generate_nature --seed 0 --task render --input_folder outputs/indoors/coarse --output_folder outputs/indoors/frames

# Use blender to extract ground-truth (optional)
python -m infinigen_examples.generate_nature --seed 0 --task render --input_folder outputs/indoors/coarse --output_folder outputs/indoors/frames -p render.render_image_func=@flat/render_image

Once complete, you can open outputs/indoors/frames and navigate to view the results.

Next Steps

See ExportingToExternalFileFormats and ExportingToSimulators to export to OBJ/USD.

We also provide an OpenGL-based ground truth extractor which offers additional ground truth channels, read more about using our ground truth here.

Generating scenes automatically

To generate a single scene in one command, you can run the following:

python -m infinigen.datagen.manage_jobs --output_folder outputs/my_dataset --num_scenes 1 --pipeline_configs local_256GB.gin monocular.gin blender_gt.gin indoor_background_configs.gin --configs singleroom.gin --pipeline_overrides get_cmd.driver_script='infinigen_examples.generate_indoors' manage_datagen_jobs.num_concurrent=16 --overrides compose_indoors.restrict_single_supported_roomtype=True

To create a large dataset of many random rooms, we recommend:

python -m infinigen.datagen.manage_jobs --output_folder outputs/my_dataset --num_scenes 1000 --pipeline_configs local_256GB.gin monocular.gin blender_gt.gin indoor_background_configs.gin --configs singleroom.gin --pipeline_overrides get_cmd.driver_script='infinigen_examples.generate_indoors' manage_datagen_jobs.num_concurrent=16 --overrides compose_indoors.restrict_single_supported_roomtype=True

You can inspect outputs/my_dataset/SEED/ to see the running logs of the subprocesses and output results.

See ConfiguringInfinigen.md for documentation on manage_jobs and commandline options.

Developer Guide

More documentation coming soon.

Restricting the solver to certain rooms / objects

Configuring compose_indoors() and ``restrict_solving()` via gin allows you to only solve sub-parts of the default constraint problem:

python -m infinigen_examples.generate_indoors --seed 0 --task coarse --output_folder outputs/indoors/coarse -g fast_solve -p \
    compose_indoors.terrain_enabled=False compose_indoors.solve_medium_enabled=False \
    restrict_solving.restrict_parent_rooms=[\"Kitchen\"] \
    restrict_solving.restrict_child_primary=[\"KitchenCounter\"] \
    restrict_solving.restrict_child_secondary=[\"Sink\"] \
    restrict_solving.solve_max_rooms=1 \
    restrict_solving.consgraph_filters=[\"counter\",\"sink\"] \
    compose_indoors.solve_steps_large=30 compose_indoors.solve_steps_small=30

Each of these commandline args demonstrates a different way in which you can restrict what work the system does:

  • compose_indoors.terrain_enabled=False compose_indoors.solve_medium_enabled=False turns off medium objects (paintings/chairs/ceilinglights) and also terrain. You can use this same pattern to disable any p.run_stage(name, func, ...) statement found in generate_indoors by specifying fcompose_indoors.{name}_enabled=False
  • restrict_solving.solve_max_rooms=1 specifies to only solve objects in 1 room of the house
  • restrict_solving.restrict_parent_rooms=[\"Kitchen\"] specifies to only solve objects in kitchen rooms. You can see infinigen/core/tags.py for available options.
  • restrict_solving.restrict_child_primary=[\"KitchenCounter\"] specifies that when placing objects directly onto the room, we will only consider placing KitchenCounter objects, not other types of objects. You can see infinigen/core/tags.py for available options.
  • restrict_solving.restrict_child_secondary=[\"Sink\"] specifies that when placing objects onto other objects, we will only consider placing Sink objects, not other types of objects. You can see infinigen/core/tags.py for available options.
  • restrict_solving.consgraph_filters=[\"counter\",\"sink\"] says to throw out any constraints or score_terms keys from home_furniture_constraints() that do not contain counter or sink as substrings, producing a simpler constraint graph.
  • compose_indoors.solve_steps_large=30 compose_indoors.solve_steps_small=30 says to spend fewer optimization steps on large/small objects. You can also do the same for medium. These values override the defaults provided in fast_solve.gin and infinigen_examples/configs_indoor/base.gin

These settings are intended for debugging or for generating tailored datasets. If you want more granular control over what assets are used for what purposes, please customize infinigen_examples/indoor_asset_semantics.py which defines this mapping.

If you are using the commands from Creating large datasets you will instead add these configs as --overrides to the end of your command, rather than -p

Generate Rooms with Floating Objects

To enable floating objects in a room, add the override compose_indoors.floating_objs_enabled=True. For example:

python -m infinigen_examples.generate_indoors --seed 0 --task coarse --output_folder outputs/indoors/coarse -g fast_solve.gin singleroom.gin -p compose_indoors.floating_objs_enabled=True compose_indoors.terrain_enabled=False restrict_solving.restrict_parent_rooms=\[\"DiningRoom\"\]

By default, between 15 and 25 objects are generated and have their size normalized to fit within a 0.5m cube. The number of objects can be configured with compose_indoors.num_floating and normalization can be disabled with compose_indoors.norm_floating=False. Collisions/intersections between floating objects and existing solved objects are off by default and can be enabled with compose_indoors.enable_collision_floating=True and compose_indoors.enable_collision_solved=True.

Run unit tests

pytest tests/ --disable-warnings