Naver Labs Wins Two Top Spots at ECCV 2024 with Spatial Intelligence Tech

Milan, Italy - Naver Labs, the research arm of South Korean tech giant Naver, clinched first place in two categories at the 2024 European Conference on Computer Vision (ECCV), showcasing its prowess in spatial intelligence technology.

The company's AI tool, MASt3R, outperformed entries from Google, Apple, and Meta in the map-free visual re-localization challenge. It demonstrated accurate positioning without relying on preexisting 3D or HD maps. 

MASt3R, an upgrade of the previously developed DUSt3R, can simultaneously process thousands of large-scale image data, creating 3D models of complex environments, including building interiors and urban areas.

Naver Labs also secured the top position in the BOP (Benchmark for 6D Object Pose Estimation) challenge for the second consecutive year. 

Its model was recognized as the best for RGB-only object positioning and fastest processing, accurately estimating the 3D rotation and position of objects in images.

Both winning technologies are based on CroCo, a 3D vision foundation model developed by Naver Labs Europe, which was acquired in 2019.

Dr. Sangok Seok, CEO of Naver Labs, emphasized the company's commitment to advancing spatial intelligence technology, stating it will be crucial for future developments in robotics, autonomous driving, and digital twin applications.

Naver Labs has consistently performed well in global tech competitions, winning first place at CVPR 2019 with its VL (Visual Localization) technology, 'R2D2'.

In a related development, Naver Cloud had 11 AI research papers accepted at ECCV 2024, focusing on multimodal AI and computer vision advancements. 

Notable research includes methods for effectively refining large amounts of image and text data for AI model production and techniques for improving image recognition model performance.

Naver plans to leverage these visual AI capabilities to enhance its generative AI services, including the recently launched HyperCLOVA X. 

The company claims that HyperCLOVA X's image understanding ability reaches 99.94% of top global models based on eight benchmarks, including ChartQA, DocVQA, MathVista, and MMMU.