Great idea, just a bit too many labels, suffers from overfitting
#4
by
GrandChieftain
- opened
Was looking for an improvement of segformer-b2-clothes because they just designate every top as upper clothes, this model ended up splitting up the same hoodie/sweater into far too many labels, rather take broadly accurate than specifically inaccurate but I appreciate the idea
Thanks you. My approach has been to always club the predictions that are of my interest to form a full segment mask. Let's say I need a segment mask of upper clothes, we can club all mask predictions for labels belonging to the upper clothes. The only one label that will cause issue is pockets. I plan to improve the model but working on a dataset is resource and time intensive. Thank you for taking out time to post this.