ex2333/scan2bim_data_5mm

Scan2BIM Point Clouds

This dataset contains point clouds associated with 81 building-floor entries. The data is organized as 46 WebDataset TAR shards. Each shard contains one point-cloud sample and its JSON metadata. A sample may cover multiple floors; shared point clouds are stored once. Point coordinates and attributes are preserved in their existing LAS/LAZ representation.

Files

• data/scan2bim-XXXXX.tar: WebDataset shards.
• manifest.json: sample keys, shard paths, file sizes, point counts, and floor associations.
• floors.json: all 81 floor entries and their corresponding samples.
• checksums.json: SHA-256 hashes and sizes of uploaded shards, added when upload completes.
• completion.json: upload completion summary, added after all shards are verified.

Each TAR contains adjacent members with a common sample key:

scene00000.las   # or scene00000.laz
scene00000.json

Metadata contains the sample key, filename, byte size, point count, and floor associations. The split field in each floor association preserves the existing train/validation assignment. Some samples cover floor entries from both splits; filter at the floor level when constructing an evaluation dataset. The archives contain point-cloud files and metadata; annotation arrays are not included.

Reading

Install webdataset to iterate over samples. LAS/LAZ members are returned as bytes and can be passed to a point-cloud reader such as laspy (with a LAZ backend for compressed files).

import json
import webdataset as wds

dataset = wds.WebDataset("downloaded/data/scan2bim-{00000..00045}.tar", shardshuffle=False)
for sample in dataset:
    metadata = json.loads(sample["json"])
    pointcloud_bytes = sample.get("las", sample.get("laz"))
    print(metadata["key"], metadata["floors"], len(pointcloud_bytes))

Full samples can be large. For memory-efficient extraction, use Python's tarfile module and copy the LAS/LAZ member to disk in chunks instead of loading the entire sample into memory. Download shards with the Hugging Face Hub client using your authorized account when access approval is required.

Uploading is incremental. Use completion.json to confirm that all 46 shards are present before consuming the complete dataset.