Research map / Detection and segmentation / Tracking
InFlux: A Benchmark for Self-Calibration of Dynamic Intrinsics of Video Cameras
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces InFlux, a real-world benchmark of 386 videos (143K+ frames) with per-frame ground-truth camera intrinsics obtained from lens-metadata lookup tables built with a modified Kalibr, and shows current intrinsics predictors struggle.
Problem
Camera intrinsics often change within in-the-wild videos (zoom, focus), but existing benchmarks either keep intrinsics fixed or lack per-frame ground truth, so dynamic intrinsics prediction cannot be evaluated.
Approach
Record per-frame lens metadata (lens focal length and focus distance) from zoom lenses and map it to intrinsics via per-lens lookup tables built from board- and drone-based calibration, with interpolation between calibrated points; Kalibr is modified for better convergence and accuracy.
Key results
- GeoCalib performs best among baselines but only 52.9% of frame-point pairs have EPE under 300 px (table reports 52.9; text says 54.1%).
- COLMAP fails to produce intrinsics for 92% of input frames, and its CFL predictions are wildly inaccurate (about 1270% fx error).
- DroidCalib fails to predict intrinsics for 15% of input frames; the modified Kalibr converges reliably on synthetic experiments where original Kalibr fails.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GeoCalib instead of WildCamera for intrinsics prediction | InFlux baseline evaluation · InFlux | % EPE < 300 px: 47.2 -> 52.9 | Helped |
| Use GeoCalib instead of WildCamera for intrinsics prediction | InFlux baseline evaluation · InFlux | % fx Error: 45.6 -> 56.5 | Mixed |
| Use GeoCalib instead of WildCamera for intrinsics prediction | InFlux baseline evaluation · InFlux | % cx Error: 5.04 -> 9.88e-2 | Helped |
Surprising findings
- Per-frame methods like GeoCalib still give high EPE errors, and per-video methods like COLMAP fail on most frames, showing current intrinsics predictors are not reliable for dynamic intrinsics.
Limitations
- Ground truth relies on per-lens lookup tables with interpolation between calibrated LFL-FD points, which may introduce error at uncalibrated settings.
- Only two zoom lenses on one camera body (ARRI Alexa Mini) are used, limiting hardware diversity.
Experimental setup
Datasets: InFlux (self-collected calibration and benchmark videos)
Benchmarks: InFlux
Metrics: percent error of fx/fy percent error of cx/cy percent of frame-point pairs with EPE < 300 px
Baselines: GeoCalib WildCamera UniDepthV2 DroidCalib Perspective Fields COLMAP
Builds on: Kalibr COLMAP Brown-Conrady distortion model
Closest papers by meaning
- TAPVid-3D: A Benchmark for Tracking Any Point in 3D (NeurIPS 2024 · similarity 0.87)
- AnyCam: Learning to Recover Camera Poses and Intrinsics from Casual Videos (CVPR 2025 · similarity 0.86)
- Reconstruct, Inpaint, Test-Time Finetune: Dynamic Novel-view Synthesis from Monocular Videos (NeurIPS 2025 · similarity 0.86)
- TrackCraft3R: Repurposing Video Diffusion Transformers for Dense 3D Tracking (NeurIPS 2026 · similarity 0.86)
- LAVIB: A Large-scale Video Interpolation Benchmark (NeurIPS 2024 · similarity 0.86)
- AC3D: Analyzing and Improving 3D Camera Control in Video Diffusion Transformers (CVPR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.