Roei Herzig

Hi there! I'm Roei, a first-year CS Ph.D. student at Tel Aviv University, working with Prof. Amir Globerson and Prof. Trevor Darrell, and a member of the Berkeley AI Research Lab.

I'm also a Machine Learning & Deep Learning Researcher, I have worked at Nexar and Trax Image Recognition in the last 5 years. Previously, I graduated magna cum laude from Tel Aviv University with MSc (CS), BSc (CS) and BSc (Physics).

I'm looking for a strong MSc students that wish to collaborate and publish in top-tier conferences on Geometry and Learning for 3D and Video Understanding.

Email  /  Twitter  /  Github  /  LinkedIn  /  Google Scholar

I mainly focus on machine learning models and deep learning methods for a structured semantic understanding in videos and images (e.g. Structured Prediction). I believe our world is compositional and humans don't perceive the world as raw pixels. Moreover, structured models can enjoy the properties of generalization and inductive-bias, which I find critical, especially at the intersections of vision, language and robotics.

Research Interest:

• Vision & Language: Object Detection, Scene Understanding, Visual Reasoning.
• Vision & Robotics: Transfer Learning, Structured Representation, Semantic Understanding.
• Machine Learning & Deep Learning: Semi-Supervised Learning, Self-Supervised Learning, Unsupervised Learning, Generative Models, Graph Neural Networks.

I'm a proud father of Adam, and when I'm not working, I'm also a history buff and love learning about science, politics, the two World Wars, the Cold War, and music.

Compositional Video Synthesis with Action Graphs

We introduce the formalism of Action Graphs, a natural and convenient structure representing the dynamics of actions between objects over time and show we can synthesize goal-oriented videos on the CATER and Something Something datasets. Finally, we demonstrate how Action Graphs can be used for generating novel compositions of unseen actions.

Learning Canonical Representations for Scene Graph to Image Generation

We present a novel model that can inherently learn canonical graph representations and show better robustness to graph size, adversarial attacks, and semantic equivalent, thus generating superior images of complex visual scenes. We show improved performance of the model on three different benchmarks: Visual Genome, COCO and CLEVR.

We propose a novel compositional action recognition task where the training combinations of verbs and nouns do not overlap with the test set. We show the effectiveness of our approach on the proposed compositional task and a few-shot compositional setting which requires the model to generalize across both object appearance and action category.

We propose an intermediate “graph-like” representation (DSGs) that can be learned in an end-to-end manner from the supervision for a downstream visual reasoning task, which achieves a new state-of-the-art results on Referring Relationships task.

We propose a latent inter-object graph representation for activity recognition that explores the visual interaction between the objects in a self-supervised manner.

We propose a hybrid coarse-to-fine approach that leverages visual and GPS location cues with on a new large-scale driving dataset based on video and GPS data.

We collect a new SKU-110K dataset which takes detection challenges to unexplored territories, and propose a novel mechanism to learn deep overlap rates for each detection.

We propose a novel invariant graph network for mapping images to scene graphs using the permutation invariant property, which achieves a new state-of-the-art results on Visual Genome dataset.