Earlier this year, I wrote a path tracing program that randomized a 3D scene of a toddler in a crib, in order to generate synthetic training data for an computer vision model. I posted about it here<\/a>.<\/p>\n I made this for the DIY infant monitor I made for my son. My wife and I are now about to have our second kid, and consequently I decided to revisit this dataset\/model\/software and release a version 2.<\/p>\n In this version, I used Stable Diffusion and Mid Journey to generate images for training the model. These ended up being way more realistic and diverse. I paid a few hundred dollars to generate over a thousand training images and videos (useful for testing detection + tracking). I labeled them manually, with LabelMe. Right now, all images have segmentation masks, but I’m in the middle of adding bounding boxes (will add key points, after that, for pose estimation).<\/p>\n To make sure this dataset actually works in practice, I created a “reference model” to train. I used various different backbones, settling on MobileNet V3 (small) and a shallow U-Net detection head. The results were pretty good, and I’m now using it in my DIY infant monitoring system.<\/p>\n Anyway, you can find the repo here<\/a> and download the dataset, which is a flat numpy array, on Kaggle<\/a><\/p>\n Cheers!<\/p>\n PS: Just to be clear, I made this dataset, it is synthetic (GenAI), it is not a paid dataset.<\/p>\n<\/div>\n submitted by \/u\/taylorcholberton <\/a>
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
