wtf is horizon3d?

Horizon3D is the placeholder repository for a research paper accepted at ECCV 2026, written by a group of researchers at Seoul National University. The README makes clear that no code has been published here yet, only a description of the method and a link to a separate project page. The repository exists mainly to hold the eventual code release once the paper is presented. The paper itself is about helping a self-driving car see objects that are far away by combining two different kinds of sensors, a radar and a camera. Radar is good at judging distance but gives a rough, sparse picture of the world, while a camera gives rich detail but struggles to judge distance on its own, especially for things that are far from the vehicle. Horizon3D combines the two so the car can build an accurate 3D picture of what is around it, and around ahead of it, even at long range. The README describes the method as using something called Gaussian primitives together with sparse features from a bird's eye view representation of the scene, processed through three named stages abbreviated KGGI, OCSF, and DPTF. Without the actual code or a more detailed explanation, it is not possible to say exactly what each stage does. The authors report that on a benchmark called TruckScenes, their method beats the previous best approach by 3.0 points on one scoring measure and 1.6 points on another measure of detection accuracy. For now, this repository is mostly informational. Anyone wanting to actually use or reproduce this work would need to check back later for the promised code and pretrained models, or visit the linked project page for more detail in the meantime.