EE Seminar:Scene Grounding In the Wild

הרישום לסמינר יבוצע באמצעות סריקת הברקוד למודל (יש להיכנס לפני כן למודל,  לא באמצעות האפליקציה)- הרישום מסתיים ב- 15:10

Registration to the seminar will be done by scanning the barcode for the Moodle (Please enter ahead to the Moodle, NOT by application)- Registration ends at 15:10

Electrical Engineering Systems Seminar

Speaker: Tamir Cohen

M.Sc. student under the supervision of Dr. Hadar Averbuch-Elor and Prof. Shai Avidan

Sunday, 25th May 2025, at 15:00

Room 011, Kitot Building, Faculty of Engineering

Scene Grounding In the Wild

Abstract

Structure from motion techniques typically produce multiple partial 3D reconstructions when applied over Internet photo collections that capture the world's landmarks. In this work, we present a technique that grounds each partial reconstruction to a complete 3D reference model of the scene, which can be easily obtained from dense imagery acquired from Google Earth Studio. We represent the reference model using 3D Gaussian Splatting, and propose a robust optimization-based framework that predicts the 6DoF+scale alignment between the reconstructed real-world images and the reference model. Our approach is built upon the premise that Internet images and views rendered from low-quality reference models capture the same underlying semantics of the scene, even though their visual appearance highly varies. We therefore augment the 3D Gaussians with semantic features, which are used to guide our optimization procedure. To evaluate our method, we contribute the WikiEarth dataset, that registers over thirty existing partial 3D reconstructions with reference models. Our evaluation shows that our approach significantly outperforms alternative baselines, allowing for grounding partial 3D reconstructions into a global unified model. We will release our code, data and trained models.