Parenting Slop Dataset

The story behind this dataset started when I was on maternity leave. I noticed that social platforms started suggesting more movies with children to me (which isn't surprising). Among them, I got a lot of synthetically generated movies of different animals stealing or attacking babies. The more I clicked on them, the stranger the animals appeared in the clips. At some point, I started downloading them, since I thought other people wouldn't believe how strange the content was. My top find was a cassowary escaping with a little kid in its beak (I even needed to check the bird's name, because it's not popular in Poland, where I live). As a researcher, and because my last project involved working on NSFW content with children, I decided to add more diverse data to synthetic movies and create a database for further analysis of children's presence in real and AI-generated videos. However, with this story, I intended to explain the bias in the database toward clips with animals and children. The Parenting Slop dataset consists of 20 movie clips showing children in various situations. All of them were downloaded from the Facebook Reels platform. The dataset contains both real and synthetic data. The clips were selected based on the presence of minors in different age groups and settings. A minor category is understood as a child under 14 years old of apparent age. Selected videos contain subjects in close-up and from a distance; children can be in groups, occluded, moving, and not looking at the camera to make the data more challenging. All the videos were trimmed to 10 seconds to have a uniform length. They were then sampled at 5 fps and saved as JPG images. As a result, 50 frames were extracted from each video, for a total of 1000 frames. Keyframes were annotated with rectangular bounding boxes for the `minor` category using LabelMe. Adults and other objects in the images are not labeled. Bounding boxes and annotations follow the COCO object detection format.