A comparison of different perspectives for AI driven detection of cooking-steps in real time

This report studies real time detection of when a user is done with a step in a recipe with the help of multimodal language models. The purpose is to compare the performance and time of two state-of-the-art models LLaVA 1.6 and Gemini 2.5 Flash. This is done from different camera angles from both first and third person views. The data consists of 45 different cooking scenes recorded from 5 different angles. The models are simulated to run in real time by incremental access to new data. The results show that Gemini has essentially better performance then LLaVA but Gemini also have issues with consistently classifying scenes as done or not done. The conclusion is that reasoning models give better performance when detecting when a cooking step is finished.