Abstract:

In this talk, I present my thesis contributions that deal with issues arising when trying to exploit the large body of data available for computer vision tasks.  
In particular we address the problem of unsupervised object discovery in time-varying, large-scale image collections by proposing a novel tree structure that closely approximates
the Minimum Spanning Tree and present an efficient construction approach along with an incremental update mechanism of the tree structure that incorporates new data as they are added to the image database.

We then focus on defining novel 3D convolutional and recurrent operators over unstructured 3D point clouds. The goal is to learn point representations for the task of 3D semantic segmentation. We overcome the limitations of the unstructured and large-scale nature of the 3D point clouds by defining local structure through two clustering methods and expand the limited receptive field of previous approaches by modeling long-range relationships with the use of Recurrent Networks.

In the third part, we address the task of interactive object segmentation where a computer vision algorithm segments an object aided by a human user. We present a method that significantly reduces the number of required user clicks compared to previous works. We use the sparse user corrections to adapt the model parameter on-the-fly during test time. In particular, we look at out-of-domain settings where the test datasets are significantly different from the datasets used to train our deep learning model.