fod-unomaha / fod-data

A dataset of foreign object debris (FOD) designed for computer vision applications.

Foreign Object Debris in Airports (FOD-A) Dataset Overview

The FOD-A dataset consists of images of common Foreign Object Debris (FOD) with a runway or taxiway background. While the main annotation style consists of bounding boxes, FOD-A also includes seperate light-level and weather categorization annotations. FOD-A is designed to be easily expanded using a command-line tool developed alongside the dataset. The instructions for this process are contained in a pdf file included in this repository.

Download

It is recommended to use the Pascal VOC format for experimentation and the original format for dataset extension. A train and validation split is provided in the Pascal VOC version. Experiments in the original paper used Pascal VOC version 2.1 (300x300 resolution) with the splits provided in that version. Tools for resizing the annotations and converting the original format to the Pascal VOC format are included in the Tools folder of this repository.

Most Current Version Download

FOD-A Version 2.1 original format (8.3 gb) 400x400 image size.
FOD-A Version 2.1 Pascal VOC format (412 mb) 300x300 image size.

Citation

If you find this dataset beneficial to your work, consider citing the paper:

Travis Munyer, Pei-Chi Huang, Chenyu Huang, and Xin Zhong. 2021. FOD-A: A Dataset for Foreign Object Debris in Airports. https://arxiv.org/abs/2110.03072

Travis Munyer, Daniel Brinkman, Chenyu Huang, and Xin Zhong. 2021. Integrative Use of Computer Vision and Unmanned Aircraft Technologies in Public Inspection: Foreign Object Debris Image Collection. arXiv: https://arxiv.org/abs/2106.00161

fod-data's Issues

Conversion Of annotations to COCO format

We have FOD-A data sets in two resolutions

1. Original (400x400)
2. Pascal Voc (300x300)

In the original annotations are numbered with images starting from frame_000000.PNG till the number of images in that folder.
For example the class "battery" has four folders (1, 2, 2a, 3) and each folder starts with frame_000000.PNG.

However, in the Pascal VOC format , the frames are numbered as frame_000000.jpg to 033792.jpg.

I have two questions:

1. If i want to convert original format to coco format, do i need to rename images/ annotations in original format for each category to avoid repitition of image names?
2. If i select Pascal Voc format(300x300) dataset and convert its annotations, according to my understanding due to lower resolution, i should get less AP for my algorithm. Do you think it will be a bad idea to use pascal voc format ? if yes why is it available?

Thanks in advance

cant see the trian dataset

I want to get the train dataset used in this . It would be really grateful if you can upload that too . I only got the fod dataset . Thanks .

Repeating file names in test.txt

In the latest edition of Pascal VOC format (300x300), under mentioned filenames are repeated in "ImageSets/Main/test.txt"

1)014567
2)015731
3)016763
4)016653

Although It can be easily solved by deleting the repeated files manually but if remain un-noticed might give error in annotation conversion. I am attaching the corrected file after deleting all repetitions

Question about the results

The author mentions in the paper ". In this experimentation, YOLOv3 produces categorization accuracy of 12.42% and a mean IOU of 47.58% on FOD-A validation data" & "SSD provides categorization accuracy of 71.81% and a mean IOU of 68.05%."

However, using the same algorithm in my experiments obtained much higher results. Is it because of the resolution? ( I used VOC format data 300*300 )