We extend our interest in self-supervised learning that we have developed in human activity discovery. In this research, we aim to develop an ability in intelligent systems to be able to distinguish one object from another object, or to discover different objects in its environment without being trained in supervised manner, i.e. without the labels. The ultimate goal is to enable an autonomous agent such as a personal robot to autonomously discover different objects in its environment from continuous observation of its surrounding through vision sensors.
Application of Convolutional Autoencoder in Object Discovery in Images
This is Azri’s BSc final year project. In this work, we applied the Convolutional Autoencoder to train a model to encode the images of a set of objects to facilitate the clustering of the images into different objects. We are starting our work in this area with simple datasets of digits (MNIST) and alphabets (A-Z). We investigated the effectiveness of a few different feature extraction techniques and different clustering algorithms in objects discovery.
- Muhammad Adli Azri Haji Mahari
- Ong Wee Hong
- Owais Ahmed Malik