We provide the code to download images of ICCWD and evaluate three different minor detection methods on it.
The code can be run on a commodity laptop and does not require specific compute, as it queries external APIs, although 8 or more cores are preferrable for faster results from DeepSeek API.
The dataset is available on HuggingFace as this link.
To download the images, create an environment using the following commands:
conda create -n iccwd python=3.11.8
conda activate iccwd
conda install pip
pip install datasets opencv-python pandas pdqhash pillow requests tqdm
Then run the following scripts:
python dataset_download_csv.py
python dataset_download_images_verify.py --folder=dataset
To label the captions in the dataset using DeepSeek, install the following dependencies.
pip install openai scikit-learn
Create an account on the DeepSeek platform and top it up with a small amount, e.g., 1$ (we used less than 0.25$ to label all the captions, but the price varies during the day). Create an api key, download it and save it to a text file under api_key.txt.
Run the following script to label the captions:
python compute_deepseek_labels.py --num_procs=1 --batch_size=1
You can adjust --num_procs depending on the number of cores available on your machine. Increasing it will allow the script to run faster by parallelizing the requests. Alternatively, you can replace multiprocessing with multithreading.
The script outputs the TPR and FPR.
To label the images in the dataset using Amazon Rekognition Image, install the following dependencies.
pip install boto3 matplotlib
Then setup an AWS account as described here. The total cost for labeling 10,000 images is estimated at 12$.
Run the following script to label the images:
python compute_amazon_labels.py
The script outputs the TPR and FPR of minor detection using Amazon Rekognition Image and two different classification rules, and saves the ROC curve to amazon_roc_curve.pdf.
To obtain results using the combined DeepSeek API with Amazon Rekognition Image, run the following script:
python compute_combined_labels.py --amazon_rule=min-range
The script outputs the TPR and FPR of the combined method using the min-range classification rule. You can set --amazon_rule=mid-range to output results using the mid-range classification rule.
If you find this dataset useful, please cite it as follows:
@article{kireev2025manually,
title={A Manually Annotated Image-Caption Dataset for Detecting Children in the Wild},
author={Kireev, Klim and Cre{\c{t}}u, Ana-Maria and Meier, Raphael and Bargal, Sarah Adel and Redmiles, Elissa and Troncoso, Carmela},
journal={arXiv preprint arXiv:2506.10117},
year={2025}
}