Deep Multiple-Instance Learning for Stable Attribute Classification in Classroom Video

Wang, Jiani

Etd

Deep Multiple-Instance Learning for Stable Attribute Classification in Classroom Video

Public Deposited

In this thesis, we explored how to classify multiple attributes of each person in a classroom video that are stable over short periods of time, such as their gender, role (student vs. teacher), and skin tone. This can benefit the field of automatic classroom analysis by giving teachers better feedback about their teaching and about possible biases they may have towards certain students. We tackled this problem using a deep Multiple-Instance Learning (MIL) method. Our experimental results on a video dataset of real classroom videos suggest that the MIL strategy is useful for classifying the stable attributes of the people in classroom videos and can improve the accuracy especially in the binary classification tasks. In addition, the model MIL MAX always performances best for all of the tasks among all the models. Finally, data augmentation and data oversampling were helpful in our experiment for solving poor model performance problem due to data imbalance.

Creator