I’ve been doing a lot of work on building computer vision models to track infants in cribs, since becoming a parent. Recently I’ve tried to start making models and datasets that are more generalized and not just for my kid. Turns out this is pretty difficult, since there aren’t a lot of datasets made for tracking infants in cribs.<\/p>\n
I made a first attempt at producing a synthetic dataset that can be used to bootstrap a model. The idea is you’d either supplement the synthetic data with a small subset of real data, or something else like transfer learning. The dataset was made using path tracing, so it looks a little bit better than some of the other synthetic datasets on infants that I’ve seen (links on my GitHub repo).<\/p>\n
Relevant Links:<\/p>\n
https:\/\/github.com\/tay10r\/infant-detection-dataset<\/a> https:\/\/www.kaggle.com\/datasets\/tay10r\/synthetic-infant-dataset<\/a> <\/p>\n It’ll be a week or so before the full dataset is done rendering (10k images). I’m traveling over the weekend so I was only able to upload a subset of the dataset (a little over 100 images).<\/p>\n Currently I use a trained model I made with about 2000 labeled images on my kid to analyze sleep patterns. I’m hoping this dataset, perhaps after a few improvements, will help produce more general models for this type of work. I’m curious to know if anyone else finds this interesting or practical. Let me know what you think!<\/p>\n<\/div>\n submitted by \/u\/taylorcholberton <\/a>
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
