Previous Work

I’ve experimented with a couple of instance segmentation tasks for an autonomous kitchen project. One task involved segmenting individual ingredients in a mixed pile, allowing robots to pick them up one by one. The other task aimed to estimate the distribution of ingredients on a pizza for quality control purposes. I trained a Mask R-CNN model using synthetic datasets generated by the Unity engine. However, it didn’t progress beyond the proof of concept stage, as creating 3D models with enough variation for various ingredients was time-consuming. Additionally, accurately segmenting stacked and intertwined objects proved to be quite challenging.

Segment Anything from Meta

Upon the release of the Segment Anything model by Meta, I was eager to explore its potential for these applications.