README.md

---
pretty_name: SpatialEncoder native WebDataset shards
task_categories:
- object-detection
---

# SpatialEncoder WDS release (in progress)

This repository contains a partition of `spatialencoder-wds-native-v1`, released
as **uncompressed WebDataset tar shards**, normally about 1 GiB. All five
repositories are parts of the same release; consult each `manifest.json`.
The manifest lists **only uploaded shards whose remote size and SHA-256 have
been verified**. An incomplete manifest is not a complete dataset.

New uploads use bucketed paths such as
`wds_v20260913/UCO3D_shards/uco3d-train/00017/uco3d-train-017395.tar`.
Each leaf holds at most 1,000 tar/JSON pairs. Already uploaded legacy paths
remain unchanged. Use the explicit paths or pinned URLs in the manifests;
do not assume every shard sits in a single dataset directory.

## Scope and splits

All 23 sources in the current SpatialEncoder training catalog are in scope.
Original train/val/test labels are retained, including val-only HOPEImage.
ScanNetpp is regenerated from all 856 official labeled training scenes and 50
validation scenes, using all good DSLR frames (`frame_stride=1`). Official
COLMAP poses, pinhole undistorted intrinsics and mesh-derived visible boxes are
used. This is not the two-scene smoke catalog. The configured pinhole training
source does not include iPhone or equirectangular panorama imagery. Unlabeled
test scenes are not relabeled as training data. DSLR has no released dense
depth target here; the native invalid-depth mask is preserved.

## Sample layout

Each sample's entries are contiguous, with one shared basename:

```
<key>.meta.json
<key>.asset00000.jpg
<key>.asset00001.npy
...
```

JSON contains dataset, original split, camera intrinsics/poses, object
annotations, native scale/depth conventions, source identity and asset
bindings. Ordinary video blocks own 64 clip starts plus up to 14 lookahead
frames, preserving every native 15-frame window across shard boundaries.
Singleton-image sources use one frame per sample. Original image and depth
encodings are retained byte-for-byte; random training augmentation is not
frozen during export.

Kubric stores the complete native TFRecord *record* containing a sequence,
not an entire multi-record source shard. UCO3D stores a sequence MP4, HDF5
depth, camera/box metadata and its filtered frame list; official excluded
observations are retained as an explicit exclusion list. CA-1M stores native
wide RGB/depth, shape records, official registered poses and JSON-converted
annotations. Metadata never requires remote pickle execution.

## Streaming from HF or S3

`reader.py` adapts raw WDS samples to the matching SpatialEncoder native
training loader, preserving camera/depth corrections. Use it alongside the
SpatialEncoder checkout and its dependencies (WebDataset, torch, OpenCV,
Pillow, numpy, h5py, protobuf, etc.). Linux is required for its memory-backed
seekable asset handles. HDF5 additionally needs one sequence-sized temporary
file per active worker, deleted after decoding; set `SPATIAL_WDS_TMPDIR` to a
local SSD or sufficiently large tmpfs. No original source directories or
whole-shard downloads are needed.

```python
import webdataset as wds
from reader import SpatialWDSDecoder

# Replace with your uploaded S3 object(s); use an IAM role or standard AWS
# credentials outside this code. This does not require a full local download.
urls = ["pipe:aws s3 cp s3://YOUR_BUCKET/PREFIX/shard-000000.tar -"]
raw = wds.WebDataset(
    urls, shardshuffle=False,
    nodesplitter=wds.split_by_node, workersplitter=wds.split_by_worker,
)
decoder = SpatialWDSDecoder(training=True)
data = raw.map(decoder).select(lambda sample: sample is not None)
```

For HTTPS, pass shard URLs from `urls/train.txt` (authenticated HTTP access
is needed for private repos). HF shard URLs are pinned to their upload commit.
The `pipe:` command must contain only trusted bucket/key strings.

The default adapter samples one owned clip start per block visit. For an exact
clip-coverage pass, flatten `decoder.iter_block(raw_sample)` instead. Dataset
mixing weights, epochs, worker seeds and distributed batch sizing remain
training configuration choices; equal source weighting must not be inferred
from a concatenated shard list. Native filtering may return `None` for a clip
with no usable objects, just as in the original loader.

Each repo manifest includes logical source, split, byte count, sample count,
SHA-256 and pinned URL per shard. `frames` includes temporal lookahead overlap;
it must not be interpreted as a count of unique training frames.

Source datasets retain their own licenses and access terms. This format does
not grant additional redistribution rights or combine their licenses.

## Missing source files

WildDet3D and hyperism packaging skips referenced source files that return
ENOENT. Missing metadata excludes that catalog row; missing RGB/depth excludes
the whole WDS block (including lookahead). Frame positions and surviving sample
keys are never renumbered. Corrupt metadata, invalid/empty assets and permission
errors remain fatal. Other sources retain strict packaging by default; the
catalog option `skip_missing_files` can explicitly override this policy.

Each run records `production/<dataset>/missing_sources.json` and an append-only
JSONL under `missing_sources/`. Completed audit logs and summaries are published
under `filtering/<dataset>/` with manifest provenance and remote hash checks.
Counts denote skipped blocks/metadata rows, not necessarily unique frames.
This records unavailability, not proof that a quality filter removed a file.
For hyperism the unavailable test partition is omitted, never relabeled train.
An entirely missing source fails instead of publishing an empty completion.

Use the matching SpatialEncoder checkout for native training: its loader skips
missing referenced source files, replenishes from valid clips, and bounds
consecutive skips. Set `SPATIAL_MISSING_SOURCE_LOG_DIR` for per-worker JSONL
audits or `skip_missing_files=False` for strict loading. WDS uses embedded
assets and does not need the original paths; broken WDS asset bindings still
raise errors. Existing checkpoint keys/fingerprints are always validated.