Stony Brook Team Wins Top Honor for New 3D Shape Algorithm

Irene Kim, a third-year Ph.D. student in the Department of Computer Science at Stony Brook University, has received a Best Paper Honorable Mention at the 2026 European Conference on Computer Vision (ECCV) in Malmö, Sweden. Her work, titled “Poppy: Polarization-Based Plug-and-Play Guidance for Enhancing Surface Normal Estimation,” placed second among 10,473 total submissions to the international conference.

The algorithm was developed by Kim and a team of co-authors including Sai Tanmay Reddy Chakkera, Alexandros Graikos, and Dimitris Samaras. The project originated in the Photon Intelligence Lab, which is led by assistant professor Akshat Dave. The lab focuses on building visual systems that use the physics of light to interpret the physical world. The team’s goal was to improve how computers perceive the shape of objects, particularly those that are shiny, dark, or lack visible texture.

Standard computer vision methods rely on ordinary color images to determine surface normals, or the direction a surface faces. This method often fails when an object is too dark to reflect enough light, too smooth to show detail, or too reflective to distinguish its form from its surroundings. Poppy solves this by analyzing polarization, a property of light that indicates the direction of light wave oscillation. Kim explained that light reflects off surfaces in predictable ways based on material and orientation, providing a physical clue about shape that standard color data misses.

A major advantage of Poppy is that it does not require training a new model from scratch. Creating new models usually demands large datasets, but polarization data paired with accurate shape measurements is rare and difficult to collect. Instead, the team used a plug-and-play strategy inspired by earlier work from Alexandros Graikos. The method refines the predictions of existing models using polarization signals from a new scene, avoiding the need for extensive retraining or changes to the original architecture.

The approach proved effective across various models and scene types. Kim noted that she was surprised by how small adjustments to the input image, guided by polarization, could lead to meaningful improvements in estimated shape. The method also enhanced 3D reconstruction from multiple views and supported image editing tasks. Akshat Dave, the project’s lead advisor, attributed the award to the combined strengths of his lab’s focus on computational imaging and the Computer Vision Lab’s expertise in generative models.

Dave stated that the recognition encourages the team to continue their work on helping machines see better by using light properties invisible to standard cameras. He sees strong potential for the technology in augmented and virtual reality, where accurate shape data helps capture real environments. In robotics, the method could allow machines to grasp objects under difficult conditions, such as a transparent, shiny beaker in dim light. Dave added that similar refinement techniques could eventually use other light properties captured by additional sensors on robots, extending the core idea beyond polarization.