NAVER LABS’ spatial intelligence research has been accepted at ECCV 2026, one of the world’s leading computer vision conferences. The accepted papers cover 3D spatial reconstruction, object recognition for autonomous driving, and indoor 3D scene generation, presenting new ways to understand real-world spaces faster and more precisely.
NAVER LABS Europe presented BLASt3R, a follow-up to DUSt3R, which has gained significant attention in the global 3D vision community. (BLASt3R: Bundle Adjustment of Any Image Set with Multi-view Matching and Monocular Priors) BLASt3R reconstructs spaces in 3D by analyzing the relationships between multiple images. It significantly improves both the speed and accuracy of 3D reconstruction compared to existing models and incorporates SLAM, the technology that allows a robot to estimate its own position and map its surroundings as it moves. In this way, it opens up the possibility of real-time 3D reconstruction in on-device environments such as robots, without relying solely on servers.
In the field of autonomous driving, “Open-Vocabulary BEV Segmentation with 3D-Aware Geometric Constraints” was also accepted. This research enables autonomous driving systems to recognize not only objects from categories encountered during training, but also previously unseen object categories.
The research utilizes a vision-language model (VLM) to understand what objects in an image are and reflect this in a real-time Bird’s-Eye View (BEV) map. A BEV map represents the surrounding environment as if viewed from above, helping vehicles and robots understand where objects are and how they relate to one another in space. The research is significant in that it goes beyond distinguishing a fixed set of object categories, allowing systems to respond flexibly even in unfamiliar environments.
InSpace, a study on indoor scene generation, reconstructs a room’s structure and the objects inside it in 3D from just a single 360-degree panoramic image (InSpace: Structure-Aware 3D Indoor Scene Generation from a Single 360° Image). At its core, the technology identifies structures such as walls and floors from limited image information and generates the objects placed within them. Recognized for its academic excellence, the paper was selected for the ECCV Spotlight, a distinction given to only about 1-2% of all submissions.
While these studies address different fields, they share a common direction: enabling AI to go beyond simply recognizing images and comprehensively understand the structure of a space, where objects are, and how the surrounding environment changes.