InFlux++: Combining Real and Synthetic Data for Estimating Dynamic Camera Intrinsics

Summary

InFlux++ (arXiv:2607.05389) is a benchmark and dataset suite from Princeton researchers for estimating dynamic camera intrinsics from video. Camera intrinsics are critical for recovering 3D structure from 2D video, yet most 3D algorithms assume fixed intrinsics, an assumption real-world videos often violate. InFlux++ consists of two components: InFlux++ Synth, a large-scale procedurally generated synthetic video dataset with over 441,000 annotated frames and 1,841 high-resolution videos providing accurate per-frame ground-truth intrinsics, and InFlux++ Real, an expanded real-world benchmark adding over 514,000 frames and 334 high-resolution videos. Fine-tuning existing intrinsics prediction methods on InFlux++ Synth improves focal length estimation accuracy on both real-world benchmarks, demonstrating that synthetic supervision is a promising approach for RGB-based camera intrinsics prediction. Project website: https://influx.cs.princeton.edu/