Cooking Object's State Identification Without Using Pretrained Model

TL;DR

This paper presents a CNN trained from scratch to identify object states in robotic cooking, achieving 65.8% accuracy without relying on pretrained models, addressing a less-explored but crucial aspect of robotic manipulation.

Contribution

The paper introduces a novel CNN architecture trained from scratch for cooking object state recognition, bypassing pretrained models.

Findings

Achieved 65.8% accuracy on unseen test data.

Demonstrated effectiveness of scratch-trained CNN in cooking state recognition.

Evaluated model performance from multiple perspectives.

Abstract

Recently, Robotic Cooking has been a very promising field. To execute a recipe, a robot has to recognize different objects and their states. Contrary to object recognition, state identification has not been explored that much. But it is very important because different recipe might require different state of an object. Moreover, robotic grasping depends on the state. Pretrained model usually perform very well in this type of tests. Our challenge was to handle this problem without using any pretrained model. In this paper, we have proposed a CNN and trained it from scratch. The model is trained and tested on the dataset from cooking state recognition challenge. We have also evaluated the performance of our network from various perspective. Our model achieves 65.8% accuracy on the unseen test dataset.