Bridge Dataset

This dataset is composed of 500 images each containing at least one bridge. This dataset has samples collected from different regions around the world, which increases its diversity and representativeness due to the varying image and bridge properties, such as the orientation of the construction, landscape context (river or mountain), area (urban or countryside), etc.
All images have 4,800×2,843 pixels and were manually annotated by our team. The bridges were first selected based on the OpenStreetMap (a publicly available annotation mapping framework) and then exported from Google Earth.

Statistics

As mentioned, the whole dataset has 500 images each containing at least one bridge. Images have 4800×2843 pixels and were manually annotated by our team.

More statistics about the dataset next:

Please, cite this dataset as:

Keiller Nogueira, Caio da Silva, Pedro Gama, Gabriel Machado and Jefersson Alex Dos Santos. “A Tool for Bridge Detection in Major Infrastructure Works Using Satellite Images.” 2019 Workshop of Computer Vision (WVC) IEEE, 2019.

Fashion Dataset

This dataset is a composition of fashion images and associated tags and comments crawled from two fashion-related social networks, namely pose.com and chictopia.com. The first part of this dataset (related to the pose.com website) was crawled from January 15, 2014 to January 25, 2014, resulting in more than two thousand instances. Now, the second part of this dataset (related to the chictopia.com website) was crawled from January 25, 2014 to February 5, 2014, resulting in almost two thousand samples. At the end, the whole dataset is composed of approximately four thousands images and associated tags and comments.

Statistics

As mentioned, the whole dataset has approximately four thousands instances.
Combining labels from both websites results in a set of 31 class possibilities: “bag”, “bathing suit”, “belt”, “booties”, “cape”, “coat”, “dress”, “glass”, “gloves”, “hat”, “headband”, “jacket”, “jewelry”, “jumpsuit”, “pants”, “pumps”, “sandals”, “scarf”, “shirt”, “shoes”, “shorts”, “skirt”, “sneakers”, “socks”, “suit”, “sweater”, “tights”, “umbrella”, “underwear”, “vest” and “wallet”. Images crawled from pose.com have 620×620 pixels while the ones crawled from chictopia.com have 400×600 pixels.

More statistics about the dataset next:

Please, cite this dataset as:

Nogueira, Keiller, Adriano Alonso Veloso, and Jefersson Alex Dos Santos. “Pointwise and pairwise clothing annotation: combining features from social media.” Multimedia Tools and Applications 75.7 (2016): 4083-4113.

Region-based Annotated Child Pornography Dataset (RCPD)

This dataset is a private database that belongs to the Brazilian Federal Police. The paper “A Benchmark Methodology for Child Pornography Detection” describes the structure of the dataset. The aim of the dataset is to assess and compare the performance of child pornography detection methods. Its files are not available to the public by any means, but researchers can submit their child pornography detection methods to be evaluated against the RCPD dataset following the instructions below.

Evaluation Instructions

To submit your method, you should send an email to macedo.jjmn@dpf.gov.br with a zip file containing your method and a script for a docker container, defining libraries and dependencies. Alternatively, you can email a link for these files. You should also implement the code to test your method, with the following requirements:

• assume that the images are located in a folder named “data”, under your method’s directory
• output a csv file named results.txt, containing the name of each tested image and your prediction (‘True’ or ‘False’) for child pornography:

filename001.jpg,True
filename002.jpg,False
filename003.jpg,False
    

The following script shows an example of this code in Python:

import os
import random
from os.path import abspath, dirname, isfile, join

# predicts whether an image is related to child pornography
# returns True or False
def predict(file_path):
    #call your method here
    return random.choice([True, False])

if __name__== "__main__":
    script_dir = dirname(abspath(__file__))
    dset = join(script_dir, "data")
    results = join(script_dir, "results.txt")
    output = open(results, "w")
    fnames = [f for f in os.listdir(dset) if isfile(join(dset, f))]
    for f in fnames:
        res = predict(join(dset, f))
        output.write(f + "," + str(res) + "\n")
    output.close()
    

Scoreboard

Results of methods and tools evaluated against the RCPD dataset:

Please, cite this dataset as:

Macedo, Joao, Filipe Costa, and Jefersson A. dos Santos. “A Benchmark Methodology for Child Pornography Detection.” 2018 31st IEEE SIBGRAPI Conference on Graphics, Patterns and Images (SIBGRAPI). 2018.