iiCEMAN's picture
Add PACO-LVIS (v2 unified release, annotations only; images from COCO 2017)
b0d869f verified
|
Raw History Blame Contribute Delete
8.1 kB
---
language:
- en
license: other
license_name: source-specific-terms
license_link: https://huggingface.co/datasets/COT5415/COT5415_Computer_Vision/blob/main/docs/data/SOURCES_AND_LICENSES.md
task_categories:
- image-segmentation
tags:
- computer-vision
- hierarchical-segmentation
- part-whole
- partimagenet
- pascal-part
- paco
- course-project
size_categories:
- 10K<n<100K
---
# COT5415 Computer Vision: unified part-whole data
Prepared by **Wugeng Zheng and Rafid Ahmed** for the UCF Computer Vision project *Seq2Seq Models for Hierarchical Image Segmentation*.
Three part-whole datasets, **91,163 images** in total, converted into one record format: one JSON line per image, listing every whole and part as a node with a compressed COCO RLE mask and a `parent_id` link.
| Source | Train | Val | Test | Images | Wholes | Parts | Images in this repo? |
| --- | ---: | ---: | ---: | ---: | ---: | ---: | --- |
| PartImageNet OOD | 16,540 | 2,957 | 4,598 | 24,095 | 24,067 | 111,960 | yes |
| Pascal-Part (processed subset) | 4,653 | 4,772 | — | 9,425 | 21,682 | 58,537 | yes |
| PACO-LVIS v1 | 45,790 | 2,410 | 9,443 | 57,643 | 273,952 | 502,054 | **no: download from COCO** |
| **Total** | | | | **91,163** | **319,701** | **672,551** | |
## Quick start
```bash
pip install huggingface_hub numpy pillow pycocotools tqdm
hf auth login # accept the access request on this page first (auto-approved)
hf download COT5415/COT5415_Computer_Vision --repo-type dataset --local-dir cv_data
cd cv_data
unzip datasets/release/unified_partwhole_v2.zip -d . # -> datasets/unified/<source>/<split>.jsonl
python code/tools/prepare_data.py # PartImageNet + Pascal-Part images into place
python code/tools/fetch_coco_images.py --splits val # PACO-LVIS images from COCO (omit --splits for all, ~9 GB)
```
Annotations only (about 0.4 GB instead of 4.6 GB):
```bash
hf download COT5415/COT5415_Computer_Vision --repo-type dataset --local-dir cv_data \
--include "datasets/release/*" "code/*" "docs/*" "README.md"
```
Read records in Python, from inside `cv_data`:
```python
import sys; sys.path.insert(0, 'code')
from partwhole import UnifiedDataset
ds = UnifiedDataset('datasets/unified/PACO-LVIS/val.jsonl')
record = ds.record(0) # annotations only; no image needed
sample = ds[0] # {'record', 'image' (PIL), 'masks': {node_id: bool array}, 'invalid_node_ids'}
```
Windows users can run the same commands in PowerShell, using `Expand-Archive` instead of `unzip`.
## Files
| Path | Contents |
| --- | --- |
| `datasets/release/unified_partwhole_v2.zip` | **Current release:** all three sources, schema, category tables, five-level grouping, quality reports, code and docs, with an embedded SHA256 manifest. Annotations only. |
| `datasets/release/unified_partwhole_v1.zip` | Earlier release (2026-11-01): PartImageNet and Pascal-Part only. Kept unchanged. |
| `datasets/archives/PartImageNet_OOD.zip`, `Pascal_Part.zip` | Original source images and labels for those two datasets |
| `datasets/archives/paco_lvis_v1.zip` | Original PACO-LVIS annotations (MIT), needed only to re-run conversion or validation |
| `datasets/archives/Pascal_Part_original_trainval.tar.gz` | Original named Pascal-Part annotations, evidence for the class names |
| `datasets/PartImageNet/images/test/n01494475/n01494475_742.JPEG` | The one test image missing from the OOD archive, recovered from the authors' Seg archive |
| `code/` | `partwhole.py` (convert, validate, load), `tools/prepare_data.py`, `tools/fetch_coco_images.py`, `tools/five_level_hierarchy.py`, `examples/read_unified.py`, `requirements-data.txt` |
| `docs/` | `UNIFIED_SCHEMA.md` and `data/` (dataset card, licenses, labels, quality policy) |
| `results/` | Example figures and validation reports |
| `MANIFEST.json` | Size and SHA256 of every file in this upload |
## Record format
```json
{"schema_version": "1.0", "dataset": "PACO-LVIS", "split": "val", "image_id": "61",
"image_path": "datasets/COCO/train2017/000000000061.jpg", "width": 640, "height": 488,
"hierarchy_semantics": "instance_object_part_links_from_source_obj_ann_id",
"nodes": [
{"id": "object:469887", "parent_id": null, "level": 0, "kind": "whole",
"category": "PACO-LVIS/whole/hat", "area": 70, "mask": {"size": [488, 640], "counts": "..."}},
{"id": "part:1496641", "parent_id": "object:469887", "level": 1, "kind": "part",
"category": "PACO-LVIS/part/hat:rim", "area": 23, "mask": {"size": [488, 640], "counts": "..."}}
]}
```
Image paths are relative to the download folder. Categories carry the dataset name, so labels from different sources are never mixed by accident. Each node also keeps its original annotation ID and label, plus quality flags.
## What "whole" and "part" mean per source
| Source | Whole (level 0) | Part (level 1) |
| --- | --- | --- |
| PartImageNet | All parts of one supercategory in the image, merged (e.g. Car) | Each original part annotation (e.g. Car Tier) |
| Pascal-Part | All pixels of one object class (e.g. person) | Each part class inside it (e.g. person/head) |
| PACO-LVIS | Each individual object instance | Each part of that instance, linked by the source `obj_ann_id` |
Only PACO-LVIS separates individual objects; in the other two sources, two people in one image share one "person" node.
## Five-level view
`code/tools/five_level_hierarchy.py render` draws each image at five levels: L1 fine part → L2 part group → L3 object → L4 super-class (animal / person / vehicle / household object) → L5 foreground. **L1 and L3 are annotated. L2, L4 and L5 are rule-based groupings of label names**, listed for every category in `datasets/unified/five_level_taxonomy.json`. They add no new pixel annotation.
## Quality notes
Conversion was checked record by record against the original annotations: IDs, categories, parent links, masks, and pixel-exact reconstruction for Pascal. Problems in the source data are kept and flagged, never repaired silently.
- **All 91,163 records passed conversion-integrity checks** with zero conversion issues. Reports: `results/preparation/validation_report.json` (PartImageNet, Pascal-Part) and `validation_report_PACO-LVIS.json`.
- **PartImageNet:** 168 malformed source polygons are kept with flags; the loader leaves those masks out. Six images have EXIF rotation, and loading follows display orientation.
- **Pascal-Part:** whole background = 16 and part background = 54. The processed labels place some parts under the wrong object class (512 train / 547 val images); these are flagged in `pascal_semantic_quality.json`.
- **PACO-LVIS:** parts and objects were outlined independently, so 79% of parts extend slightly beyond their object's mask (5.5% of all part pixels). These pixels are kept, and each part records them in `outside_parent_area`. 123 masks are empty in the source and are kept. LVIS labels are not exhaustive: an absent category is only a confirmed negative if listed in `neg_category_ids`.
- PACO-LVIS validation was run without the COCO train2017 images present, so the image-size check covered only the 2,183 val2017 images; all of them passed. Run `python code/partwhole.py validate --datasets PACO-LVIS` after downloading the images to check the rest.
## Sources and terms
- PartImageNet OOD and processed Pascal-Part: [turkeyju/PartImageNet](https://huggingface.co/datasets/turkeyju/PartImageNet), revision `cfa2edf3cf5ffbede2bd1fc934edca0aba99f042`. Named Pascal-Part annotations: [authors' page](https://roozbehm.info/pascal-parts/pascal-parts.html).
- PACO-LVIS: [facebookresearch/paco](https://github.com/facebookresearch/paco) (MIT), annotations from `https://dl.fbaipublicfiles.com/paco/annotations/paco_lvis_v1.zip`. Images are COCO 2017 ([terms of use](https://cocodataset.org/#termsofuse)): Flickr photos under their individual licenses, recorded per image in `source_image_metadata`. They are not re-hosted here.
No single license covers everything here; each source keeps its own terms. See [SOURCES_AND_LICENSES.md](docs/data/SOURCES_AND_LICENSES.md).