Naver Labs tops in spatial intelligence tech at ECCV

Naver Labs' spatial intelligence technology was recognized as superior to that of Google, Apple, and Meta at the world's leading computer vision conference.

Naver Corp. said on Tuesday that Naver Labs secured first place in two categories at the European Conference on Computer Vision (ECCV).

The ECCV is a specialized academic conference in the field of computer vision, focused on the latest advancements in artificial intelligence (AI) for images, videos, and related topics, and serves as a platform for introducing new technological trends.

At ECCV 2024, Naver Labs competed in two challenges—"map-free visual re-localization" and the "benchmark for 6D object pose estimation"—and won first place in both.

The map-free visual re-localization challenge tests the ability without precise maps.

Visual re-localization typically relies on pre-generated 3D or HD maps.

In situations like disasters or construction sites where maps cannot be generated or pre-prepared without maps is becoming increasingly important.

In this challenge, Naver Labs showcased its AI tool MASt3R, which reconstructs images into 3D. MASt3R was recognized for its ability to localize accurately even without precise maps, taking first place among 12 participating teams, including Google, Apple, and Meta.

MASt3R is an improved version of DUSt3R, which is based on CroCo, a 3D vision foundation model (VFM) developed by Naver Labs Europe.

The benchmark for the 6D object pose estimation challenge measures how accurately an AI model can estimate the 3D rotation and position of objects in an image.

Naver Labs' model, submitted for this category, was recognized for being the most accurate and fastest in determining object positions using only RGB images, securing first place.

This is the second consecutive year Naver Labs has won this benchmark, and the model is also based on CroCo.

Meanwhile, Naver Cloud was achieved at ECCV 2024, with 11 AI research papers accepted.

These papers introduced methods for refining large amounts of image and text data to build high-quality training datasets more effectively, as well as techniques for improving the performance of image recognition models.

Write to Ji-Eun Jeong at jeong@hankyung.com