KAS2003/xfield-nlos-paired

XField paired data

This repository contains train,validation only. 199 chunks, 3890 complete present/absent frame pairs. Layout: //chunk_NNN/. Root manifest.jsonl lists chunk path, split, town, complete pair/frame counts, and every included file's size and SHA256. Verify all manifest files after download; CHUNK_DONE proves source processing and alone does not prove upload completion.

Town02 is validation. Training towns are under train. Town15 is stored in a separate public repository.

Format and exclusions

Processed chunk bytes are preserved: frames.jsonl, mothers.npy, native.tar, camera.tar, teacher_.npz, census.jsonl, cached rgb arrays/index, obs.json, ego_speed.json, records_clean.jsonl, CHUNK_DONE and RAW_ARCHIVE.json when present. native.tar and camera.tar are required processed shards and are included. raw_frames.tar, nlos_raw_archive and temporary raw capture folders are excluded. Source data was not deleted or rewritten. Local staging uses file links; uploaded content is the actual file bytes.

Absolute paths and download relocation

Source metadata retains absolute paths: frames.jsonl mother (mothers.npy::index) and labels/cam shard references must be rebased after download. Raw provenance fields frame_dir, raw rgb, mother_src and RAW_ARCHIVE.json archive locations describe excluded source archives and may stay unavailable. RQSD uses the included cached rgb arrays/index, mothers.npy, processed shards, teacher labels and census. Provenance paths are not substitutes for processed inputs.

The manifest records absolute_path_markers per file, a byte scan for literal /root/ and /Users/. Compressed payloads are not decoded by that scan. The producer puts keys/arrays in NPZ/NPY and file references in JSON. All original bytes, including two historically rebased migrated chunks, remain unchanged.

This repository is public under user authorization U20. Download it into an absolute DATASET_DIR; authentication is optional. Verify manifest sizes and SHA256 BEFORE relocation. Run the following only on the downloaded copy, never the source:

python3 - "$DATASET_DIR" <<'PY'
from pathlib import Path
import sys,json
root=Path(sys.argv[1]).expanduser().resolve()
for c in root.glob('*/*/chunk_*'):
    if not c.is_dir():continue
    marker='/'+c.parent.name+'/chunks/'+c.name+'/'
    def fix(v):
        if isinstance(v,dict):return {k:fix(x) for k,x in v.items()}
        if isinstance(v,list):return [fix(x) for x in v]
        if isinstance(v,str) and v.startswith('/') and marker in v:
            rel=v.split(marker,1)[1];base,sep,idx=rel.partition('::');dest=c/base
            if dest.is_file():return str(dest)+(sep+idx if sep else '')
        return v
    for f in c.rglob('*'):
        if not f.is_file() or f.suffix not in ('.json','.jsonl') or f.name=='RAW_ARCHIVE.json':continue
        s=f.read_text()
        if f.suffix=='.jsonl':out=''.join(json.dumps(fix(json.loads(l)),ensure_ascii=False)+'\n' for l in s.splitlines() if l)
        else:out=json.dumps(fix(json.loads(s)),ensure_ascii=False,indent=2)+'\n'
        f.write_text(out)
print('Downloaded metadata relocated; original manifest hashes describe the unmodified download.')
PY

The command rebases existing processed files under their recorded capture/chunk prefix and retains ::index and tar offsets. Unavailable raw provenance is left unchanged. Metadata hashes differ after local relocation by design. Set RQSD_NEWDATA to colon-separated downloaded CHUNK_DONE chunk directories; preserve train/validation definitions. Never add the sealed-test repository to that training list.