Tran1312/MSCOCO2017

MS COCO 2017 — Roboflow v2 Raw

An object detection dataset repackaged from Microsoft COCO 2017, using the Roboflow v2 raw export dated December 15, 2020, 20:55 GMT. It includes JPEG images, COCO JSON annotations, and JSONL tables for data loading, training, and analysis.

This is a repackaged dataset, not an official release by the COCO Consortium, Microsoft, or Roboflow, and does not imply their endorsement. The immediate source is documented in README.roboflow.txt, README.dataset.txt, and the info field of the COCO JSON files. See Roboflow and COCO.

License scope: COCO annotations are licensed under CC BY 4.0; image rights require verification for each image. The license: unknown metadata reflects incomplete verification of licensing for the full image collection and does not negate the annotation license. See Licensing and attribution.

Dataset size

The counts below were obtained directly from the image directories and annotations in this dataset, rather than inferred from the official COCO dataset size.

Split	JPEG images	Bounding boxes	Images without annotations	Boxes with width or height ≤ 0
train	116,408	836,322	1,002	157
val	5,000	36,335	48	18
Total	121,408	872,657	1,050	175

• There are 80 object classes, with English class names.
• This package does not include a test split.
• Filename checks confirmed that every image listed in COCO JSON has a corresponding file, with no extra images in either split directory. No annotations reference image IDs or category IDs missing from their split.
• These checks validate structure and references; they do not cover decoding every image or comparing bytes against the original COCO files.
• Do not assume this export is fully equivalent to the official COCO benchmark.

Directory structure

.
├── README.md
├── README.dataset.txt             # Bundled source information and JSONL examples
├── README.roboflow.txt            # Roboflow export information
├── images/
│   ├── train/*.jpg
│   └── val/*.jpg
├── labels/
│   ├── train/
│   │   ├── _annotations.coco.json
│   │   ├── annotations.jsonl
│   │   ├── images_info.jsonl
│   │   ├── categories.jsonl
│   │   ├── images_train.jsonl
│   │   └── class_sampling.jsonl
│   └── val/
│       ├── _annotations.coco.json
│       ├── annotations.jsonl
│       ├── images_info.jsonl
│       ├── categories.jsonl
│       └── images_val.jsonl
├── analysis_outputMSCOCO/         # Training split reports: JSON, CSV, PNG charts
├── analyst.py                    # Analysis script with locally configured paths
├── cache/                        # Local .pkl indexes, not required to load the data
├── images.tar.xz
├── labels.tar.xz
└── analysis_outputMSCOCO.tar.xz

The bundled .tar.xz files are archives; the counts in this README were verified against the extracted directories. The full archive contents have not been verified to match the current directories. When publishing, choose either directories or archives to avoid storing duplicate data, and retain both source README files. cache/ does not need to be uploaded to the Hub.

Annotation formats

COCO JSON

Each labels/<split>/_annotations.coco.json is a JSON object containing info, licenses, categories, images, and annotations.

Component	Main fields	Description
images	id, file_name, width, height, license, date_captured	Image metadata; the path is images/<split>/<file_name>
annotations	id, image_id, category_id, bbox, area, segmentation, iscrowd	One record per object
categories	id, name, supercategory	Class mapping for the Roboflow export
licenses	id, name, url	License declarations supplied by the export; these do not replace source verification for each image

bbox uses [x, y, width, height] in pixels, with the origin at the top-left corner of the image. These are neither normalized coordinates nor [xmin, ymin, xmax, ymax].

An actual example from the training split:

{"id": 0, "image_id": 0, "category_id": 16, "bbox": [0, 77, 284, 452], "area": 128368, "segmentation": [], "iscrowd": 0}

In this export, all segmentation fields are empty and all iscrowd values are 0. The package therefore provides bounding boxes, but no segmentation masks, captions, or keypoints. IDs are scoped to each split and should not be assumed to be original COCO IDs. The exported date_captured field should not be used as evidence of the actual capture date.

JSONL

Each line is a standalone JSON object.

File	Fields	Notes
images_info.jsonl	id, file_name, height, width	Minimal metadata, joined to annotations through id ↔ image_id
annotations.jsonl	id, bbox, image_id, category_id, iscrowd, isfake, area, isreflected, flag_reflected	Object annotations with zero-based class IDs
categories.jsonl	name, id	80 classes, with IDs from 0 to 79
images_train.jsonl / images_val.jsonl	image_name, path	Image list; path is a filename relative to the split image directory
class_sampling.jsonl	id, probability	80 records in the training split; probability values are not normalized probabilities in [0, 1]; the formula used to generate them is undocumented

The JSONL record corresponding to the annotation above:

{"id": 0, "bbox": [0, 77, 284, 452], "image_id": 0, "category_id": 15, "iscrowd": 0, "isfake": 0, "area": 0, "isreflected": 0, "flag_reflected": 0}

Do not mix category IDs between formats:

• COCO JSON has 81 entries in categories: entry 0 is coco-objects, and the 80 object classes use IDs 1–80.
• JSONL omits coco-objects and shifts the 80 classes to IDs 0–79.
• All annotations in both splits were compared: category_id_jsonl = category_id_coco_json - 1; the corresponding id, image_id, and bbox values match.
• For example, person has ID 49 in COCO JSON and 48 in JSONL; broccoli has IDs 16 and 15, respectively.
• Use the class table bundled with the format you are reading. Do not use the official COCO ID table directly to interpret Roboflow IDs.

All JSONL area values are 0, so this field should not be used to group objects by size. To obtain bounding box area, compute width * height after handling invalid boxes; this is not mask area.

Loading the data

This example uses the Python standard library. Run it from the dataset root after downloading or extracting the data:

import json
from collections import defaultdict
from pathlib import Path

root = Path(".")
split = "val"
data = json.loads((root / "labels" / split / "_annotations.coco.json").read_text())
categories = {c["id"]: c["name"] for c in data["categories"]}
annotations = defaultdict(list)
for ann in data["annotations"]:
    annotations[ann["image_id"]].append(ann)

item = data["images"][0]
image_path = root / "images" / split / item["file_name"]
objects = annotations[item["id"]]
print(image_path)
print([(categories[a["category_id"]], a["bbox"]) for a in objects])

This README includes YAML metadata for a Hugging Face Dataset Card. The current layout stores images and annotation tables separately and does not define a unified sample table for Hugging Face Datasets. Do not assume load_dataset(repo_id) will automatically join bounding boxes to the correct images. To display object annotations in the Dataset Viewer, prepare the appropriate metadata or convert to a supported format; see the ImageFolder guide.

Transformations and known limitations

• According to README.roboflow.txt, no image augmentation was applied during export; the preprocessing section lists no specific transformations. The image processing history has not been verified beyond this information.
• Observed differences include .rf.* suffixes in image filenames, the images/ and labels/ layout, JSONL tables, remapped class IDs, additional fields, and area=0 in JSONL. No conversion script or complete history is available to establish who made each change or when.
• There are 175 boxes with nonpositive width or height. Users need to decide whether to filter or correct them before training; this README does not modify annotations.
• analysis_outputMSCOCO/summary.json describes only the training split and uses JSONL IDs; it does not represent the full dataset. The report also records 55 boxes extending beyond image boundaries in the training split.
• The dataset has class imbalance: the training report records 252,916 person annotations and 196 hair drier annotations. Training results should be evaluated by class and intended use.
• Duplicate image content across splits and equivalence to the official COCO splits have not been checked. Do not claim official benchmark results using this export without verifying equivalence.
• analyst.py contains absolute paths from a previous environment; update its configuration variables before running it again.

Licensing and attribution

Annotations

Under the COCO Terms of Use (official source), annotations belong to the COCO Consortium and are licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). The Roboflow export also lists CC BY 4.0 in its metadata and source README.

When sharing or modifying material covered by CC BY 4.0, provide appropriate attribution, retain supplied notices, link to the license, and indicate changes; do not imply endorsement by the licensor. See the full license text.

Images and redistribution rights

The COCO Consortium does not own the image copyrights; the COCO terms require compliance with the terms of use for images from Flickr. Do not assume the annotation license applies to all images. See Flickr Creative Commons.

Image records in the current COCO JSON reference only a single license=1 entry labeled CC BY 4.0 by the Roboflow export; they do not include flickr_url, coco_url, or original photographer information. JSONL also omits the license field. The available metadata is therefore insufficient to verify redistribution rights and attribution for each image.

Before publishing the image package on Hugging Face, match images against the original COCO/Flickr metadata, record the creator, source URL, and applicable license, and comply with the terms for each image. If verification is incomplete, publish only the permitted annotations with instructions for obtaining images from official sources, rather than treating this README as permission to redistribute images.

Citing the paper does not replace licensing obligations or guarantee protection from copyright claims. This package claims no ownership of the original images or annotations.

Dataset attribution

Underlying dataset: COCO Consortium and the Microsoft COCO authors. Export source: Roboflow, Microsoft COCO 2017, v2 raw, exported on December 15, 2020. The JSONL tables and reorganized layout are described in this README. Image copyrights belong to their respective owners; attribution for each image must be added after source verification.

Citation

When using this dataset, cite the COCO paper and acknowledge the Roboflow export. The first BibTeX entry below references the arXiv version of the paper, not a new DOI for this package.

@misc{lin2014microsoftcoco,
  title         = {Microsoft COCO: Common Objects in Context},
  author        = {Tsung-Yi Lin and Michael Maire and Serge Belongie and
                   Lubomir Bourdev and Ross Girshick and James Hays and
                   Pietro Perona and Deva Ramanan and C. Lawrence Zitnick and
                   Piotr Doll{\'a}r},
  year          = {2014},
  eprint        = {1405.0312},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/1405.0312}
}

@misc{roboflow2020cocov2raw,
  author       = {{Roboflow}},
  title        = {Microsoft COCO 2017: v2 Raw Export},
  year         = {2020},
  howpublished = {Roboflow Public Datasets},
  url          = {https://public.roboflow.com/object-detection/microsoft-coco-subset},
  note         = {Export timestamp: 2020-12-15 20:55 GMT, recorded in the bundled README.roboflow.txt}
}

For reproducibility, also record the Hugging Face repository URL and the exact commit/revision used once the dataset is published.

References

1. Lin, T.-Y. et al. Microsoft COCO: Common Objects in Context. arXiv:1405.0312.
2. COCO Consortium. Dataset and downloads; Terms of Use.
3. Roboflow. Microsoft COCO 2017 Object Detection Dataset. Local version information: README.roboflow.txt.
4. Creative Commons. CC BY 4.0, Legal Code.
5. Flickr. Creative Commons and image license types.
6. Hugging Face. Dataset Cards, Repository Licenses, ImageFolder.