Data

Two separate things live here. data/maps holds the floor plans, which are inputs. data/coverage holds the generated training set, which the collectors write and training reads. They are separate roots because a coverage chunk is not a floor plan and the two are usually on different disks.

The test, out-of-distribution and real maps are at Maps, in its maps/ folder. The 18 training maps and the coverage dataset (about 1 TB) are not public yet; collect_data regenerates the dataset from the training maps. Three of the maps are included, one of each kind.

Map	Kind
data/maps/test/small_office3.pgm	a held-out KTH floor plan
data/maps/ood/aws_maps/hospital.pgm	an out-of-distribution AWS RoboMaker world
data/maps/real/ecss_corridor.pgm	the real corridor, from a SLAM run

Point either root elsewhere with an environment variable and change nothing in the source.

export INTELX_MAPS_ROOT=/path/to/your/maps
export INTELX_DATASET_ROOT=/path/to/your/coverage/data

Maps

Occupancy maps are PGM files at 0.05 m per pixel, sealed so that walls are watertight. intelx.paths.map_path resolves every one of them under data/maps, so a path in the code is always relative.

data/maps/
  train/                       the 18 training maps, 17 KTH floor plans and
                               the AWS small house
  test/                        the three held-out KTH maps
  ood/
    house_expo/{small,medium,large}/
    msd/{small,medium,large}/  Modified Swiss Dwellings
    aws_maps/                  AWS RoboMaker worlds
  real/                        the real corridor scan

Sources

The maps are standardised from four public datasets, each under its own licence: KTH floor plans, HouseExpo, Modified Swiss Dwellings and the AWS RoboMaker worlds. The maps/ folder of the download unpacks into the tree above.

data/train_metadata.csv lists the 18 training maps with the shows each receives, Table 1 of the paper, allocated in proportion to the complexity score of Eq. (1). intelx.cli.collect_data reads it.

The Coverage Dataset

intelx.paths.dataset_path resolves everything the collectors write and everything training reads, under data/coverage.

data/coverage/
  coverage_data_main_3/    what intelx.cli.collect_data writes, per map
  train_chunks/            the chunked form intelx.cli.train reads
  dense_data/<map>/        what a single-map run of intelx.data.driver writes

It is not public yet, so generate it. python -m intelx.cli.collect_data runs the collection across every map in INTELX_MAPS_ROOT, and it is a long multi-GPU job, so read the constants at the top of that file first. <CHUNKING_SCRIPT> then packs what it writes into train_chunks/.