Blame for evaluation/README.md

# PartProbe3D frozen evaluation

Historical MessyKitchens mask archives now reside under [_deprecated/messykitchen/original_masks/evaluation](https://huggingface.co/datasets/cubec/PartProbe3D-preprocessed/tree/main/_deprecated/messykitchen/original_masks/evaluation).
Current evaluation downloads exclude these archives.

This bundle contains 2,603 single-image evaluation samples and their GT object meshes.
The PartProbe3D benchmark requires this dataset.
The Hugging Face repository stores this dataset under `evaluation/`.
The archive index records its transfer size, including RGB images, masks, previews, and shared geometry.
It contains no training samples, depth maps, model weights, or image features.

| Dataset | Images | Unique GT exports | Selection |
| --- | ---: | ---: | --- |
| PICASSO | 100 | 100 | All released images from ten scenes |
| MessyKitchens | 1,503 | 100 | All released images from 100 scenes |
| 3D-FRONT | 1,000 | 1,000 | Random prepared test views with seed 42, without replacement |

The 3D-FRONT population contains 10,066 prepared views from 935 test rooms.
The selected views cover 635 rooms and retain their test assignments.
Each MessyKitchens scene has one GT export, shared across its views through relative manifest paths.
All archive members contain regular files.

## Mandatory files

These paths refer to the Hugging Face repository root.
The complete benchmark requires `evaluation/index.json` and every archive that it lists.

| File or files | Count | Contents |
| --- | ---: | --- |
| `evaluation/index.json` | 1 | Archive inventory and SHA-256 checksums |
| `evaluation/metadata.tar` | 1 | Portable manifests, original selections, and source provenance |
| `evaluation/picasso-00000.tar` | 1 | All 100 PICASSO images, masks, previews, and GT meshes |
| `evaluation/messykitchen-<manifest-hash>-*.tar` | See index | All 1,503 MessyKitchens images, reviewed masks, independent channels, registration records, audit flags, previews, and 100 shared GT meshes |
| `evaluation/3dfront-00023.tar` and `3dfront-00024.tar` | 2 | The selected 1,000 3D-FRONT test images, masks, previews, and GT meshes |

The downloader restores every archive and checks its checksum.
Each view requires its RGB image, instance mask, and referenced GT mesh.
Older project checkouts also require `code/evaluation.patch` before restoration.

The project README stays outside the data archives.
The Hub stores it separately at `evaluation/README.md`.
Restoration preserves the local project README, including local edits.
Metadata archives retain dataset provenance files such as `messykitchen/SOURCE_README.md`.

## Optional files

These separate Hub components are optional for this benchmark:

| File or directory | Purpose |
| --- | --- |
| `optional/held_out/<dataset>/` | Remaining test observations beyond the training monitors |
| `optional/geometry/<dataset>/` | Source geometry for new targets or other evaluation sets |
| `optional/occlusion/<dataset>/` | GT relations for applicable SS configurations |
| `optional/evaluation/ycbv/` | Additional YCB-V evaluation data |

The mandatory bundle already includes the GT meshes for all 2,603 images.
Its restoration needs no optional component or original source dataset.
Training files under `training/` serve model training separately.

## Restore

Use the current PartProbe3D checkout for restoration:

```bash
python -m pip install -e '.[preprocess]'
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --task eval
```

The downloader pins one Hub revision and verifies every archive checksum before restoration.
The required evaluation bundle restores to `datasets/evaluation/`.
Default training downloads exclude this bundle.
The source manifest at `code/evaluation_manifest.json` records the loader files and patch checksum.

## Refresh Hub metadata

Use this procedure to remove the project README from an earlier archive.
The refresh command downloads only the evaluation index and metadata archive.
It preserves every other member and every data archive record.
It publishes the project README separately.
It requires no transfer of image, mask, or mesh archives.

Prepare the update without uploading:

```bash
python -m preprocess.evaluation refresh-metadata \
  --output outputs/huggingface/evaluation_metadata
```

Publish the three prepared metadata files:

```bash
python -m preprocess.evaluation refresh-metadata \
  --output outputs/huggingface/evaluation_metadata --upload
```

The publication uses a pinned parent commit.
A concurrent Hub change stops the update.
Archive checksums and existing data checks remain mandatory.

Refresh the task catalog after publication:

```bash
python -m preprocess.tasks publish --repo cubec/PartProbe3D-preprocessed \
  --output outputs/huggingface/tasks
```

The catalog records the current Git revision.
Use a revision available on the Git remote.

## Geometry-only inference

The benchmark compares geometry meshes.
Comparison baselines disable independent appearance decoding, UV generation, and texture baking.
Geometry checks must show unchanged surfaces, object placement, indices, units, and coordinate frames.
Native geometry settings remain fixed.

Save raw prediction GLBs, stable sample identities, the pinned dataset revision, and scoring settings.
Keep source snapshots, dependency versions, and metric results for portable reevaluation.
Restore RGB, masks, and GT meshes from the pinned Hub release when required.
Geometry metric reevaluation needs no model weights or native working files.

## Read samples

```python
from dataloader.frozen import FrozenEvaluationLoader

dataset = FrozenEvaluationLoader('datasets/evaluation/messykitchen')
sample = dataset[0]
rgb, masks, meshes = sample.rgb, sample.masks, sample.gt_meshes
```

Use `--dataset frozen` for shared evaluation commands.
The configuration files are `evaluation/configs/picasso_frozen.json`, `evaluation/configs/messykitchen_frozen.json`, and `evaluation/configs/front3d_frozen.json`.
Original source datasets are unnecessary for this loader.

```bash
python -m evaluation.evaluate select --dataset frozen \
  --dataset-config evaluation/configs/messykitchen_frozen.json \
  --count 1503 --seed 42 --workers 4 --output outputs/evaluation/messykitchen_frozen
```

Each dataset contains `manifest.json`, `source_selection.json`, `views/`, and `ground_truth/`.
The manifest records file checksums, original sample indices, mask settings, and dataset provenance.
RGB pixels and mask labels preserve their frozen selection checksums.
GT files preserve the original exported GLB bytes and their float32 vertex representation.
The original geometry hash remains provenance under `source_gt_sha256`.

PICASSO and 3D-FRONT GT meshes use OpenCV camera coordinates in meters.
MessyKitchens GT meshes use the registered scene coordinates in meters.
Its reviewed SAM3 masks use GT-assisted correspondence and camera recovery.
The label images resolve overlapping channels by native SAM3 quality, with smaller object IDs breaking ties.
Each view also retains the exact independent channels and registration records.
All 13 flagged views remain in the 1,503-view set.
The masks have no independent pixel annotations or released photo calibration.
Use the `default` protocol (one global Sim(3)) for MessyKitchens scoring.

The source loader uses `observations/registered/manifest.json` instead of the historical SAM3 observations.
The portable loader reads the same labels from `datasets/evaluation/messykitchen`.
Historical Hub revisions preserve the original masks and archive checksums.
Existing predictions and selections retain their historical inputs.
Create a separate selection before evaluating this mask release.

Restoration refuses to replace files with different checksums.
Before upgrading an existing installation, preserve its MessyKitchens directory and cohort index under `datasets/_deprecated/messykitchen/`.
Then restore the requested datasets with the current task command.
The [release report](https://github.com/cube-c/PartProbe3D/blob/main/reports/messykitchen_registered_release_20261024.md) records the published revision and migration paths.

## Sources and terms

Original dataset terms continue to apply to their images and meshes.
This derived bundle introduces no replacement license for source data.

- [PICASSO official release](https://github.com/MIT-SPARK/Picasso#data) supplies RGB images, instance annotations, object meshes, and poses.
- [MessyKitchens official project](https://github.com/MessyKitchens/MessyKitchens) supplies the real photographs and registered scene meshes under CC BY-NC 4.0.
- [3D-FRONT official project](https://tianchi.aliyun.com/specials/promotion/alibaba-3d-scene-dataset) supplies the rooms and furniture used for our rendered test views.

MessyKitchens authors are Junaid Ahmed Ansari, Ran Ding, Fabio Pizzati, and Ivan Laptev.
Its original dataset card and citation remain in `messykitchen/SOURCE_README.md`.
PartProbe3D adds predicted SAM3 masks for MessyKitchens and prepares the 3D-FRONT captures.