hab3_bench_assets / README.md
aclegg3
Upload episode datasets and generator config + README changes to document.
c6b323f
|
raw
history blame
2.36 kB
metadata
license: cc-by-nc-4.0

Habitat v0.3.x Benchmark Dataset

Assets, configs, and episodes for reproduceable benchmarking on Habitat v0.3.x.

Setup

Clone this repo and symblink as data/hab3_bench_assets in habitat-lab directory. Download the Habitat compatable YCB SceneDataset and create a symbolic link in data/objects/ycb or use the habitat-sim datasets_download script (README).

Contents:

  • Scene Dataset: hab3-hssd/ - the necessary configs and assets to load a subset of HSSD dataset into habitat-lab and utilize it for Hab3 rearrangement tasks.

  • Episode Datasets: episode_datasets - a set of serialized RearrangeDataset files generated for the benchmark SceneDataset. See "Generating New Episodes" below for details.

  • hab3_bench_ep_gen_config.yaml - config file for generating new RearrangeDataset files.

Generating New Episodes:

The provided config hab3_bench_ep_gen_config.yaml is available for generating new hab3 benchmarking episodes. It defines the scene, objects, and generator configs (e.g. number of clutter objects).

The generator command should be run on a Habitat 3.0 compatable branch (e.g. SIRo) with the included assets from fpss/fphab commit cd1549303d759abacbb377a8dd52c5f7af0d0e5a as follows:

python -u habitat-lab/habitat/datasets/rearrange/run_episode_generator.py --config data/hab3_bench_assets/hab3_bench_ep_gen_config.yaml --run --verbose --num-episodes 10 --seed 0 --out data/hab3_bench_assets/episode_datasets/large_large.json.gz

Naming of the episode file <scene_complexity>_<object_complexity>.json.gz depends on the following parameters:

Scene Complexity:

Currently we are testing on 3 differently sized scenes:

  • small: 103997919_171031233 (area 35.92) - 1 bed, 1 bath
  • medium: 108736635_177263256 (area 55.49) - 3 bed, 2 bath
  • large: 102816009 (area 172.43) 4 bed, 4 bath + den & office

One of these scene sets must be selected in the config before generation.

Object Complexity:

Currently we are testing 3 clutter object size complexities:

  • small: 2 objects
  • medium: 5 objects
  • large: 10 objects

One of these sampler params must be selected in the config before generation.