yangqinzhe/LEVIRDet-159-Dataset

LEVIRDet-159

A Million-Scale 159-Category Dataset and Foundation Model for Universal Remote Sensing Object Detection

Paper · GitHub · Project page · Interactive demo

LEVIRDet-159 is a remote sensing object detection dataset with 159 hierarchical category names, 174,497 image files, and 2,563,973 bounding-box annotations in its hierarchical annotation set. It supports coarse and fine-grained detection, hierarchical evaluation, and research on generalization across remote sensing imagery. The paper describes the dataset and its companion model, LEVIRDetNet, and reports approximately 700,000 fine-grained annotations.

This repository distributes two encodings of the same images, split assignment, and annotation geometry. Download one version. Both use the public 159-class taxonomy defined by taxonomy_159.json.

Models and Docker: The LEVIRDetNet model repository provides the 30-class and strict 159-class detector checkpoints, the ready-to-use Docker image, a matching source snapshot, and inference instructions.

Stream the original-image dataset

The original-image WebDataset contains all 174,488 annotated images: 124,257 train and 50,231 test, in 130 TAR shards. It preserves original encoded image bytes and both independent annotation sets, including all special fields and public provenance. No image is recompressed or resized. The original ZIP files and all their download links remain available below.

import json
from datasets import Features, Image, Value, load_dataset

root = "hf://datasets/yangqinzhe/LEVIRDet-159-Dataset/web_tool/webdataset_original/webdataset"
features = Features({"image": Image(), "metadata.txt": Value("string"),
                     "__key__": Value("string"), "__url__": Value("string")})
data = load_dataset("webdataset", features=features,
                    data_files={"train": f"{root}/train-*.tar",
                                "test": f"{root}/test-*.tar"},
                    split="train", streaming=True)
sample = next(iter(data))
image = sample["image"]
metadata = json.loads(sample["metadata.txt"])
annotations_30 = metadata["annotations_30"]
annotations_159 = metadata["annotations_159"]

The 159-class source annotations retain a coarse category_id; it is not the 159-class model output index. Resolve fine labels from the published fine name and hierarchy fields, using the ordered taxonomy_159.json vocabulary. See the field interpretation guide for the exact priority and the distinction between parent vehicle and leaf car.

Resource	Purpose
web_tool/webdataset_original/webdataset/*.tar	Complete annotated original-image release, streamed on demand
web_tool/webdataset_pilot/webdataset/*.tar	Fixed 256-image validation subset
visual-preview (default Hub configuration)	Small original/30-class/159-class annotated thumbnail gallery; not training data

Use streaming=True to avoid downloading and caching the entire dataset first. Use the explicit "webdataset" loader and Image features shown above for TARs.

Full checks: manifest, integrity audit, COCO root metadata and taxonomy, nine unannotated test files, conversion and validation tools.

Streaming and visual preview pilot

Select visual-preview in the Dataset Viewer to browse 32 examples (16 train, 16 test), with original-image thumbnails beside 30-class and 159-class annotated thumbnails.

The WebDataset pilot contains 256 original images (128 train, 128 test). Load its TAR files explicitly; the repository's default configuration loads the small Parquet gallery above, not the TAR dataset:

import json
from datasets import Features, Image, Value, load_dataset

root = "hf://datasets/yangqinzhe/LEVIRDet-159-Dataset/web_tool/webdataset_pilot/webdataset"
features = Features({"image": Image(), "metadata.txt": Value("string"),
                     "__key__": Value("string"), "__url__": Value("string")})
data = load_dataset("webdataset", features=features,
                    data_files={"train": f"{root}/train-*.tar",
                                "test": f"{root}/test-*.tar"},
                    split="train", streaming=True)
sample = next(iter(data))
image = sample["image"]
metadata = json.loads(sample["metadata.txt"])
annotations_30 = metadata["annotations_30"]
annotations_159 = metadata["annotations_159"]

Each TAR sample pairs <key>.image (unchanged original encoded image bytes) with <key>.metadata.txt (UTF-8 JSON containing both independent image and annotation records, all special fields and public provenance). The declared Image feature detects JPEG/PNG/TIFF/BMP from the original bytes. Root COCO metadata and the taxonomy are under web_tool/webdataset_pilot/source_metadata/.

Model output indices, COCO category IDs, and fine-category fields are distinct. Crowd values are retained.

See pilot validation, sample selection, and gallery validation. Existing MMDetection training commands continue to use the extracted COCO release; streaming training needs an appropriate data-loader adapter.

Contents

• Choose a release
• Dataset statistics
• Download and extract
• Extracted files
• Annotation format and taxonomy
• Evaluation notes
• Provenance and limitations
• License
• Citation
• Contact

Choose a release

Version	Images	Download size	ZIP files	Extracted directory
Original-image release	Original image bytes, with standardized filenames	111.41 GB / 103.76 GiB	62	02_release/
Lightweight release	JPEG quality 95, with original dimensions	42.32 GB / 39.42 GiB	22	03_lite_jpeg95/

Sizes include the archives and their companion files. GB is decimal; GiB is binary. The original-image version contains 46 training ZIPs, 15 test ZIPs, and metadata.zip; the lightweight version contains 14 training ZIPs, 7 test ZIPs, and metadata.zip.

The Original-image release retains source file contents and formats, including JPEG, PNG, TIFF, and BMP. The Lightweight release uses lossy JPEG encoding (quality=95, 4:2:0 chroma subsampling, optimized progressive encoding). Image dimensions and box coordinates are unchanged, but pixel values can differ.

Dataset statistics

Item	Train	Test	Total
Image files	124,257	50,240	174,497
Images listed in annotation JSONs	124,257	50,231	174,488
Bounding boxes in the 159-class annotation set	1,741,681	822,292	2,563,973
Bounding boxes in the historical 30-class annotation set	1,741,694	822,298	2,563,992

The original validation split is distributed here as test. The 30-class and 159-class annotation sets refer to the same image filenames within a split, but have slightly different object counts. They are separate annotation sets: do not join annotations by array position or assume their COCO IDs are interchangeable. The two downloadable image encodings are alternative versions of the same dataset, not additional training examples.

Download and extract

The original download releases contain independent ZIP archives and COCO-style JSON annotations. Use the download-and-extract workflow below for those complete releases. The full streaming files and the explicitly labelled pilot/preview subsets above are additional formats; they do not replace these archives.

1. Download one version

Install the client in a Python 3.10+ environment:

python -m pip install --upgrade huggingface_hub

Run the following Python code to download the lightweight version:

from huggingface_hub import snapshot_download

release = "Lightweight release"  # or "Original-image release"
snapshot_download(
    repo_id="yangqinzhe/LEVIRDet-159-Dataset",
    repo_type="dataset",
    allow_patterns=[f"{release}/*"],
    local_dir="downloads",
)

Files will appear under downloads/Lightweight release/ or downloads/Original-image release/. Keep the download cache and rerun the same command to resume an interrupted download.

For test images only, use this call instead:

from huggingface_hub import snapshot_download

release = "Lightweight release"  # or "Original-image release"
snapshot_download(
    repo_id="yangqinzhe/LEVIRDet-159-Dataset",
    repo_type="dataset",
    allow_patterns=[f"{release}/*"],
    ignore_patterns=[f"{release}/train-*.zip"],
    local_dir="downloads",
)

Always keep metadata.zip, archives.json, SHA256SUMS.txt, consumer_extract.py, and the download notes together with the selected image ZIPs. metadata.zip contains the annotation files, taxonomy, provenance mapping, and detailed field documentation.

2. Verify and extract

For a complete lightweight download, run this single-line command in a terminal:

python "downloads/Lightweight release/consumer_extract.py" --archive-dir "downloads/Lightweight release" --out-dir "datasets"

For the original-image version:

python "downloads/Original-image release/consumer_extract.py" --archive-dir "downloads/Original-image release" --out-dir "datasets"

For a test-only lightweight download:

python "downloads/Lightweight release/consumer_extract.py" --archive-dir "downloads/Lightweight release" --out-dir "datasets-test" --split test

The commands also work in Windows PowerShell with Python on PATH. Keep the quotes around paths containing spaces. The extraction script itself needs Python 3.9+ and uses only the standard library.

The script checks archive sizes and SHA-256 hashes, checks ZIP entries, and verifies CRCs while extracting. ZIPs are independent archives: do not concatenate them. --out-dir is the parent of the extracted version directory, so do not append 02_release or 03_lite_jpeg95 again. Reserve disk space for both the downloaded archives and an extracted copy.

Extraction does not overwrite or merge into an existing version directory. An interrupted run may leave a .partial directory; it is not a completed dataset and extraction does not automatically resume it. Resolve that incomplete output or use a fresh --out-dir before retrying. To change from test-only to a complete extraction, use a new output location. Test-only extraction still includes the metadata for both splits, but only test images.

Extracted files

datasets/
└── 02_release/                   # or 03_lite_jpeg95/
    ├── train/images/
    ├── test/images/
    ├── annotations/
    │   ├── train_159.json
    │   ├── test_159.json
    │   ├── train_30.json
    │   └── test_30.json
    ├── taxonomy_159.json
    ├── provenance.json
    ├── metadata.json
    ├── release_summary.json
    ├── alpha_exceptions.json
    ├── SHA256SUMS.files.txt
    ├── README_数据与标注说明.md
    └── README_下载与解压.md

Image filenames use shared seven-digit identifiers across the two versions. Training filenames span 0000001–0124257, and test filenames span 0124258–0174497. Extensions follow the original format in the original-image version and are .jpg in the lightweight version. These filename identifiers are not COCO image IDs.

Resolve an image path using the selected version's train/images/ or test/images/ directory plus the JSON image record's file_name. Do not resolve it relative to annotations/.

Annotation format and taxonomy

COCO-style records with additional hierarchy fields

All annotation files contain images, annotations, and categories. Bounding boxes use pixel coordinates [x, y, width, height], where (x, y) is the top-left corner. image_id links an annotation to an image record; category_id links it to a coarse category in that JSON.

The *_159.json files still contain 30 top-level categories records. Their 159-class supervision is carried by additional fields on each annotation, together with the shared taxonomy. Counting the top-level categories array does not give the model's hierarchical output dimension.

Field	Meaning
fine_category_name	Preferred effective label at the annotated level of detail
fine_category_level1 … fine_category_level4	Available hierarchy path; deeper levels may be absent or null
fine_supercategory	Preserved superclass information from the source annotation
original_fine_category_name	Provenance remark on three training records; never an extra training label

The following is an excerpt of one annotation, with unrelated fields omitted:

{
  "id": 1,
  "image_id": 1,
  "category_id": 10,
  "bbox": [730, 646, 22, 116],
  "area": 2552,
  "iscrowd": 0,
  "fine_category_name": "bulkCarrier",
  "fine_supercategory": "ship",
  "fine_category_level1": "ship",
  "fine_category_level2": "civil_ship",
  "fine_category_level3": "bulkCarrier",
  "fine_category_level4": null
}

Its effective label is bulkCarrier, with path ship → civil_ship → bulkCarrier. Objects can be annotated at different hierarchy depths; an unspecified child category must not be invented.

The authoritative 159-class vocabulary

taxonomy_159.json contains:

• classes and ordered_classes: the same ordered list of 159 unique names, including parent and fine classes.
• base_classes: the 30 root categories.
• class_hierarchy: each parent's immediate children.
• category_ids: a one-based mapping for derived flat annotations; this does not replace the original COCO category_id values.

Keep this class order consistent between the dataset reader, model head, checkpoint, and evaluator. Model class indices are zero-based positions in the ordered list.

For effective-label selection, use a nonempty fine_category_name first; otherwise use the deepest populated field from fine_category_level4 through fine_category_level1, then the coarse category linked by category_id. Validate the selected label against the taxonomy. Never use original_fine_category_name for label selection.

Evaluation notes

For hierarchical evaluation, count a child ground-truth box toward its ancestor categories as well. This expansion is performed in the evaluator; the published annotation files keep one record per annotated object. A conventional flat COCO evaluation is a different protocol and should be identified separately.

Training, testing, checkpoint adaptation, and inference instructions are maintained in the GitHub repository.

Citation

If you use LEVIRDet-159 or LEVIRDetNet, please cite the paper:

@misc{yang2026levirdetmillionscale159categorydataset,
  title={LEVIRDet: A Million-Scale 159-Category Dataset and Foundation Model for Universal Remote Sensing Object Detection},
  author={Qinzhe Yang and Dongyu Wang and Haohan Niu and Jia Xu and Zhenwei Shi and Zhengxia Zou},
  year={2026},
  eprint={2606.25312},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2606.25312}
}

Contact

Please report dataset or tooling issues through [email protected], GitHub Issues or this repository's Community tab.