# Interesting papers ## OKVIS2-X: Open Keyframe-based Visual-Inertial SLAM Configurable with Dense Depth or LiDAR, and GNSS - https://youtu.be/K8oZvbI7I58?si=sT1NK6rYN982dSFW - https://github.com/ethz-mrl/OKVIS2-X - https://arxiv.org/pdf/2510.04612 - visual, inertial, measured or learned depth, LiDAR, GNSS를 마음대로 활용할 수 있는 SLAM 시스템 - Dense volumetric map을 이용한다는게 특징. 그럼에도 submap 구조를 잘 활용해서, 9km 까지 scale이 가능함. - Camera extrinsic online calibration도 가능 <img width="1900" height="1073" alt="Image" src="https://github.com/user-attachments/assets/63f5933f-643b-46ec-9264-76a24c8d2ea9" /> <img width="1439" height="878" alt="Image" src="https://github.com/user-attachments/assets/f1076acd-b8c3-42b7-8f54-2231d92f7bb0" /> ## Open-YOLO 3D: Towards Fast and Accurate Open-Vocabulary 3D Instance Segmentation - https://github.com/aminebdj/OpenYOLO3D - https://arxiv.org/pdf/2406.02548 - SAM 이나 CLIP을 매 프레임마다 추출해서 multi-view reconstruction을 통해 3D instance segmentation을 하는 모델들이 많았음. 굉장히 무거운 연산들이라 속도가 많이 느림. - 2D object detection + 3D network 두개만으로 기존의 Sota보다 16배나 빠른 속도를 얻어낼 수 있음. <img width="2353" height="667" alt="Image" src="https://github.com/user-attachments/assets/9aa9765f-bec8-4a26-9eb7-5f42ff7f8049" /> <img width="1439" height="646" alt="Image" src="https://github.com/user-attachments/assets/9d14dec8-f819-4853-afbd-194715ab0081" /> <img width="1439" height="646" alt="Image" src="https://github.com/user-attachments/assets/eb26ec11-a68e-467d-83ff-c33f3daae793" />