Stony Brook has Clear Vision at CVPR 2020

The AI Institute Stony Brook University’s Institute for AI-Driven Discovery and Innovation and its faculty have been expanding their research in the field of computer vision, presenting revolutionary findings on an international scale. Three pieces of this inventive and pioneering research have landed at the chief international venue for computer vision: Computer Vision and Pattern Recognition (CVPR)

The leading global conference for computer vision collects the most prolific research within the field of the given year. Four AI institute faculty had a total of three papers accepted this year: Haibin Ling, Niranjan Balasubramanian, Minh Hoai Ngyuen, and Dimitris Samaras. The accepted papers naturally provide novel insights for those in academia and industry. In 2020, the conference was flooded with close to 7,000 submissions, with a selective acceptance rate of just 22%. Despite this spike in competition Stony Brook continues to lead the way with its three research papers accepted at CVPR 2020. This recognition distinguishes the university as an international powerhouse for computer vision.

“Stony Brook does not lack vision,” says AI Institute Director Steven Skiena. “In fact, we are ranked among the top ten departments nationally in computer vision. Our annual performance at CVPR and other top conferences reflects this.”

Computer-driven emotion recognition is a modern research topic which continues to develop and advance. A collaborative research project spearheaded by professors Balasubramanian, Ngyuen, and Samaras: “Learning Visual Emotion Representations From Web Data,” is a novel approach towards understanding emotional variables in not only humans, but scenes and symbols alike.

Their research introduces a novel neural network which extracts emotion: EmotionNet. The architecture and capacities of EmotionNet outperforms previous attempts at emotion recognition. Such competing networks generally work off of small datasets, and therefore produce limited results. Contrarily the extensive dataset, StockEmotion, where over one million web images, coupled with 690 emotion-related tags has trained EmotionNet, making for the most advanced network in emotion recognition.

EmotionNet serves as a functional tool for a variety of emotional analyses. The results of this network are integral to understanding human behaviors, and to evolve communication amongst machines and humans alike. Notably, the usefulness and developments that EmotionNet provides speaks to the Institute’s collaborative and research-driven prowess.

Professor Haibin Ling and his students have developed an esteemed Human-Object Interaction model, which benefits from a revolutionary cascade network architecture, accurately recognizes the human and object in the given image, and the interaction between the two entities. Professor Ling’s network proves to be more refined than previous Human-Object Interaction models. The primary difference is this revolutionary network localizes interactions between entities at the pixel-level, which provides more thorough and precise results.

This network proves amply more powerful than its predecessors because it functions at a pixel-level; as such, the success of this architecture is bound to inspire new research exploring this new domain. Due to this innovation, the model earned first place in the ICCV-2019 Person in Context Challenge.

These are just some examples of the Institute’s success from CVPR 2020. The next CVPR conference is being held virtually, June 19-25, 2021.