Most face detection repositories only support COCO format and Widerface format annotations. However. original FDDB dataset does not provide such annotations. This project helps create COCO format and Widerface format annotation files for FDDB.
We provide annotation files in COCO and Widerface format for 10-fold cross validation. You can find them in annotations/COCO/
and annotations/Widerface/
. Since there is no key point ground truth in FDDB dataset, we simply fake key points.
COCO format example:
{"info":
{"description": "FDDB in COCO format.",
"url": "",
"version": "1.1",
"contributor": "SSRSGJYD",
"date_created": "2020-11-15"},
"images": [
{"file_name": "2002/08/11/big/img_591.jpg",
"coco_url": "local",
"height": 431,
"width": 450,
"flickr_url": "local",
"id": 1},
...],
"annotations": [
{"segmentation": [],
"num_keypoints": 5,
"keypoints": [0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1],
"iscrowd": 0,
"area": 43139,
"image_id": 1,
"bbox": [180, 41, 179, 241],
"category_id": 1,
"id": 1},
...]
}
Widerface format example:
# 2002/08/11/big/img_591.jpg
180 41 179 241 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Our code works in Python 3.x environments. Install requirements by:
pip3 install -r requirements.txt
Download FDDB dataset from here and decompress the two folders into /fddb/images/
.
To create COCO format annotations:
python fddb2coco.py
To create Widerface format annotations:
python fddb2txt.py
@TechReport{fddbTech,
author = {Vidit Jain and Erik Learned-Miller},
title = {FDDB: A Benchmark for Face Detection in Unconstrained Settings},
institution = {University of Massachusetts, Amherst},
year = {2010},
number = {UM-CS-2010-009}
}