Title: Princeton365: A Comprehensive Dataset Featuring Precise Camera Pose Estimation

Abstract

This paper presents Princeton365, a large-scale, heterogeneous dataset comprising 365 videos annotated with high-precision camera poses. By implementing a novel ground truth acquisition framework that combines calibration boards with a 360-degree camera system, our work addresses the disparity between data diversity and accuracy found in existing SLAM benchmarks. The collection includes synchronized monocular and stereo RGB video streams, along with IMU data, covering indoor environments, outdoor settings, and object scanning scenarios.

Additionally, we introduce a scene scale-aware evaluation metric for SLAM systems. This new metric, which calculates errors based on optical flow induced by pose estimation inaccuracies, enables meaningful performance comparisons across different scenes. This stands in contrast to traditional measures like Average Trajectory Error (ATE), which do not facilitate such cross-scene analysis, thereby allowing researchers to better identify and analyze the failure modes of their algorithms. Furthermore, we establish a rigorous Novel View Synthesis (NVS) benchmark that extends beyond current standards by including challenging cases, such as fully non-Lambertian scenes and 360-degree camera trajectories. Researchers can access the dataset, code, videos, and submission guidelines at https://princeton365.cs.princeton.edu.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC