Radu Sibechi

A spatiotemporal deep neural network for semantic video segmentation by using consecutive frames from a video, where only the last one is labeled. Our network architecture relies on a novel interconnection of two components: a fully convolutional network to model spatial information and temporal units that are employed at intermediate levels of the convolutional network in order to propagate information through time. The corresponding paper was accepted at an autonomous driving workshop at ICCV 2019 in South Korea.

We addressed the problem of visual question answering, which requires both image and language understanding to answer a question about a given photograph. We described two models for this task: a simple bag-of-words baseline and an improved long- and short-term memory-based approach.

Neural machine translation (NMT) is an approach to machine translation that uses a large artificial neural network to predict the likelihood of a sequence of words, typically modeling entire sentences in a single integrated model.

In this project, we attempted to solve the NMT by using a model inspired by the sequence-to-sequence approach. However, instead of encoding the sentence using a recurrent neural network, we made use of positional embeddings. We displayed our findings and drew meaningful conclusions from them.

To make text recognition easier, it is often desirable to digitally flatten a document image when the physical document sheet is folded or curved.
In this project, we explored multiple convolutional neural network architecture types such as DeepLab, UNet, and PSPNet to predict a deformation field that can be used to unwrap the deformed input image. We made use of the novel loss function implemented in the scientific paper DocuNet™.