InFlux++: Benchmark for Dynamic Camera Intrinsics

• InFlux++ is a comprehensive data suite and benchmark for estimating per-frame camera intrinsics in videos with dynamic parameters, addressing challenges like zoom, focus, and lens breathing.
• It integrates a large procedurally generated synthetic dataset (InFlux++ Synth) and a real-world benchmark (InFlux++ Real) to offer diverse scenes and camera motions for enhanced calibration.
• Empirical findings show that finetuning existing models on InFlux++ improves focal length estimation, though challenges in distortion estimation and synthetic-to-real adaptation persist.

InFlux++ is a data suite and benchmark for estimating per-frame camera intrinsics in videos with dynamic camera intrinsics, that is, videos in which the intrinsic parameters vary from frame to frame rather than remaining fixed across a sequence. It was introduced to address two obstacles identified for RGB-based intrinsics prediction: training data is scarce and lacks intrinsics diversity, and existing benchmarks, including the original InFlux, have limited scene and camera motion diversity. InFlux++ therefore combines InFlux++ Synth, a large procedurally generated synthetic dataset for training, with InFlux++ Real, a large real-world benchmark for evaluation (Liang et al., 6 Jul 2026).

1. Definition and problem setting

Camera intrinsics determine the mapping from 3D camera coordinates to 2D pixel coordinates. In the pinhole model,

$$\begin{bmatrix} u \ v \ 1 \end{bmatrix} = K \begin{bmatrix} X/Z \ Y/Z \ 1 \end{bmatrix}, \qquad K = \begin{bmatrix} f_x & 0 & c_x \ 0 & f_y & c_y \ 0 & 0 & 1 \end{bmatrix}.$$

The parameters $f_x$ and $f_y$ are effective focal lengths in pixels, while $c_x$ and $c_y$ are the principal point coordinates. InFlux++ is concerned with the regime in which these quantities, together with distortion parameters in real cameras, are dynamic rather than fixed (Liang et al., 6 Jul 2026).

The motivating cases are videos in which a zoom lens changes the lens focal length (LFL), the camera changes focus through the lens-to-object distance (LTO), and the lens exhibits lens breathing. The thin-lens relation used in the work is

$$\frac{1}{\text{LFL}} = \frac{1}{\text{CFL}} + \frac{1}{\text{LTO}}, \qquad \text{CFL} = \frac{\text{LFL}\cdot\text{LTO}}{\text{LTO} - \text{LFL}},$$

where CFL is the effective camera focal length in physical units and is then converted to the pixel focal lengths $f_x$ and $f_y$ (Liang et al., 6 Jul 2026).

The fixed-intrinsics assumption used by many 3D pipelines is identified as a major limitation. The benchmark is explicitly motivated by the observation that methods such as SfM/SLAM, NeRF, Gaussian Splatting, and monocular depth or pose estimation often assume sequence-wise constant intrinsics, whereas real in-the-wild videos may contain frame-rate zoom and focus variation. A central premise of InFlux++ is therefore that estimating per-frame intrinsics from RGB images is critical for making 3D methods robust to videos with dynamic intrinsics (Liang et al., 6 Jul 2026).

2. Data suite architecture and camera parameterization

InFlux++ has two components with complementary roles: a synthetic training corpus and a real-world benchmark. Their scale and annotation structure are summarized below (Liang et al., 6 Jul 2026).

Component	Scale	Core content
InFlux++ Synth	1,841 videos, 441,840 frames	RGB, per-frame intrinsics $K$, pose, LFL, LTO, CFL; subset with depth and normals
InFlux++ Real	334 videos, 514K+ frames	RGB with per-frame calibrated intrinsics from lens metadata and LUTs

In the synthetic component, LFL and LTO are the primary time-varying physical parameters. The paper states that temporal variation in these quantities produces physically plausible variation in CFL, together with lens breathing, even when LFL is fixed. Synthetic distortion is added through a Brown–Conrady model. For a normalized point $\mathbf{x}=(x,y)$ with $f_x$0, the radial model is

$f_x$1

with optional tangential parameters $f_x$2. The corresponding radial displacement in normalized coordinates is

$f_x$3

and in pixels,

$f_x$4

The coefficients $f_x$5 and $f_x$6 are obtained by sampling desired pixel displacements at two edge control points and solving a $f_x$7 system, while $f_x$8 may be sampled from $f_x$9 (Liang et al., 6 Jul 2026).

This parameterization reflects a deliberate modeling choice. The work does not treat dynamic intrinsics merely as dynamic field of view; it represents them as a coupled consequence of zoom, focus, distortion, and defocus blur. A plausible implication is that InFlux++ is intended not only as a benchmark for focal-length regression, but also as an optical-model benchmark in which peripheral rays, distortion, and focus-dependent image formation are part of the supervision signal (Liang et al., 6 Jul 2026).

3. InFlux++ Synth

InFlux++ Synth is described as a large-scale procedurally generated synthetic video dataset with 441K+ annotated frames from 1841 high-resolution videos. In the detailed dataset statistics, this becomes 1,841 videos and 441,840 frames, with 10 s per video at 24 FPS and resolution $f_y$0. The dataset is divided into 1,040 indoor videos (249,600 frames) and 801 nature videos (192,240 frames) (Liang et al., 6 Jul 2026).

All videos provide per-frame RGB, intrinsics matrix $f_y$1, camera pose $f_y$2, LFL, LTO, and CFL. A subset of 818 videos (196,320 frames) additionally includes depth maps (blurred + pinhole) and surface normals (blurred + pinhole). Scene diversity includes indoor categories such as living rooms, bedrooms, bathrooms, kitchens, and offices, and nature categories such as forests, deserts, caves, mountains, coasts, arctic, and underwater. The scenes also include dynamic objects, including animals, falling leaves, and water, as well as non-Lambertian materials, reflections, and mirrors (Liang et al., 6 Jul 2026).

The rendering stack is built on Infinigen and Infinigen Indoors, with photorealistic rendering through Blender’s Cycles engine. Camera trajectories are produced through Infinigen’s Rapidly-Exploring Random Tree (RRT) planner, which samples collision-free positional waypoints and connects them with Bézier-interpolated camera poses. Additional filtering rejects views with too much sky or geometry that is too close, and applies a dominant surface normal check: a single large planar patch is considered dominant if it covers more than 65% of samples indoors or more than 80% in nature. Local replanning is used when orientation sampling fails at a waypoint, and free-space checks prevent the camera from passing through geometry (Liang et al., 6 Jul 2026).

The temporal dynamics of intrinsics are generated by bounded random walks. For LFL, keyframes follow

$f_y$3

with $f_y$4, clipped to $f_y$5, and smoothly interpolated with Bézier curves. For LTO, the paper notes that direct random walks tend to produce frames where everything is out of focus, so focus depth is instead parameterized by

$f_y$6

where $f_y$7 follows a bounded random walk, and $f_y$8 are estimated by ray tracing using 5th and 95th percentile visible depths from 1000 sampled rays. This ensures that portions of the scene remain in focus throughout the sequence (Liang et al., 6 Jul 2026).

The synthetic loader is itself part of the contribution. It adds on-the-fly radial distortion, optional tangential distortion, photometric augmentations such as gamma, tone curve, color jitter, noise, and JPEG artifacts, and geometric augmentations such as aspect ratio changes and cropping. The augmentations explicitly disable blur, sharpen, and downscale so as not to interfere with thin-lens blur cues. This suggests that InFlux++ Synth is designed as a training environment in which optical effects relevant to calibration are preserved rather than randomized away (Liang et al., 6 Jul 2026).

4. InFlux++ Real

InFlux++ Real is the benchmark component. It is described as a large-scale real-world benchmark that extends InFlux with 514K+ newly captured frames across 334 high-resolution videos, and in the detailed statistics as 334 videos, 514,000+ frames, resolution $f_y$9, and 23.976 FPS (Liang et al., 6 Jul 2026).

The hardware platform consists of an ARRI Alexa Mini with two cinema zoom lenses: Canon CINE-SERVO 17–120 mm (canon17) and Fujinon Premista 80–250 mm (premista80). These lenses provide /i Technology metadata, specifically per-frame lens focal length (LFL) and focus distance (FD). Ground-truth intrinsics, including distortion, are obtained from per-lens lookup tables (LUTs) that map $c_x$0 to intrinsics. The LUTs are constructed from calibration board captures at a grid of LFL and FD settings, followed by interpolation (Liang et al., 6 Jul 2026).

A major methodological change relative to the original InFlux is the introduction of a camera-moving board-based calibration for large field-spatial-footprint calibration. The calibration pattern consists of large AprilTag arrays on a rigid screen of approximately $c_x$1, installed in a multi-story lecture hall with balcony. The board is fixed while the camera is moved and rotated from multiple positions, including ground floor left and right, balcony left and right, and additional ground positions for shorter FD. According to the paper, this configuration excites all rotational axes, covers the entire field spatial footprint, and avoids the complexity and instability of drone-based calibration (Liang et al., 6 Jul 2026).

The benchmark is also intended to broaden evaluation coverage. Compared with the original InFlux, InFlux++ Real includes more urban scenes, more everyday/domestic indoor scenes, sports and recreation, moving vehicle captures, and elevated viewpoints. The parallax statistics are explicitly reported: the original InFlux contains 324 low, 35 medium, and 27 high parallax videos, whereas InFlux++ Real contains 119 low, 28 medium, and 183 high-parallax videos. The repaired canon17 lens is additionally reported to have a significant principal point offset, which broadens the range of $c_x$2 values (Liang et al., 6 Jul 2026).

Privacy protection is built into the release. Frames are processed with RetinaFace for face detection and EgoBlur for license plates, using full-image and $c_x$3 tile inference, confidence thresholding, and non-maximum suppression; detected regions are then blurred. This places the benchmark in the category of calibration datasets intended for public dissemination rather than only internal evaluation (Liang et al., 6 Jul 2026).

5. Evaluation protocol and baseline methods

The benchmark evaluates predictions using parameter-error recalls and reprojection-style endpoint error. For focal lengths and principal points, percent error is defined per frame as

$c_x$4

with analogous quantities for $c_x$5. Reported recalls are Recall@{1\%,10\%,20\%} for $c_x$6 and Recall@{0.5\%,1\%,2\%} for $c_x$7 (Liang et al., 6 Jul 2026).

The endpoint error (EPE) metric is defined by projecting visible 3D points with ground-truth and predicted intrinsics, including distortion, and computing the Euclidean image-plane displacement

$c_x$8

InFlux++ modifies the visibility filtering used for Brown–Conrady distortion. The distorted radial mapping

$c_x$9

is required to remain on its monotonic branch, with cutoff radius $c_y$0 given by the first positive root of

$c_y$1

A point is only considered visible if its distorted projection lies in image bounds and its undistorted radius satisfies $c_y$2. This correction is intended to exclude physically implausible points that would otherwise fold back into the image under non-monotonic distortion (Liang et al., 6 Jul 2026).

Because ground-truth intrinsics in the real benchmark come from LUT interpolation, the paper also introduces LUT-reliable EPE recall@T. Reliability is estimated by leave-one-out calibration at each LUT vertex: a vertex is marked reliable if leave-one-out interpolation achieves EPE recall@T $c_y$3 against measured intrinsics, and a frame contributes to EPE recall@T only when all vertices of its enclosing LUT cell are reliable at threshold $c_y$4. Main tables report thresholds 10, 50, and 300 px, while the full benchmark reports thresholds from 1 to 300 px (Liang et al., 6 Jul 2026).

The real benchmark provides a validation split with RGB, ground-truth intrinsics, and lens metadata, and a test split with RGB only. The recommended usage is to train or finetune on InFlux++ Synth, validate on InFlux++ Real validation, and submit test predictions to an evaluation server. The paper lists the following baselines: AnyCalib, GeoCalib, UniDepthV2, WildCamera, Perspective Fields, DroidCalib, and COLMAP (Liang et al., 6 Jul 2026).

Among these, the paper states that AnyCalib performs best overall on the combined InFlux and InFlux++ Real test frames. Its reported values include $c_y$5 recall@10\% = 25.2\%, $c_y$6 recall@1\% $c_y$7, and LUT-reliable EPE recall@50px = 25.1\%. The paper nevertheless emphasizes that the overall performance remains modest, which indicates the difficulty of dynamic per-frame intrinsics prediction from RGB alone (Liang et al., 6 Jul 2026).

6. Empirical findings, applications, and limitations

The central empirical finding is that finetuning existing intrinsics prediction methods on InFlux++ Synth consistently improves focal length estimation across both InFlux++ Real and InFlux. The paper demonstrates this with AnyCalib, finetuned for 15 epochs on 826 training videos (198k frames) from InFlux++ Synth. On the original InFlux, $c_y$8 recall@10\%) improves from 11.5\% to 17.5\%, and $c_y$9 recall@20\%) from 20.8\% to 34.1\%. On InFlux++ Real, $$\frac{1}{\text{LFL}} = \frac{1}{\text{CFL}} + \frac{1}{\text{LTO}}, \qquad \text{CFL} = \frac{\text{LFL}\cdot\text{LTO}}{\text{LTO} - \text{LFL}},$$0 recall@10\%) improves from 28.2\% to 34.2\%, and $$\frac{1}{\text{LFL}} = \frac{1}{\text{CFL}} + \frac{1}{\text{LTO}}, \qquad \text{CFL} = \frac{\text{LFL}\cdot\text{LTO}}{\text{LTO} - \text{LFL}},$$1 recall@20\%) from 45.7\% to 54.0\%. On the combined benchmark, $$\frac{1}{\text{LFL}} = \frac{1}{\text{CFL}} + \frac{1}{\text{LTO}}, \qquad \text{CFL} = \frac{\text{LFL}\cdot\text{LTO}}{\text{LTO} - \text{LFL}},$$2 recall@10\%) improves from 25.2\% to 31.2\%, and $$\frac{1}{\text{LFL}} = \frac{1}{\text{CFL}} + \frac{1}{\text{LTO}}, \qquad \text{CFL} = \frac{\text{LFL}\cdot\text{LTO}}{\text{LTO} - \text{LFL}},$$3 recall@20\%) from 41.3\% to 50.4\% (Liang et al., 6 Jul 2026).

At the same time, the paper reports that EPE recall generally decreases after this finetuning. On the combined benchmark, EPE recall@50px decreases from 25.1\% to 18.7\%, and EPE recall@10px from 2.70\% to 1.93\%. The explanation offered is that AnyCalib is trained with an $$\frac{1}{\text{LFL}} = \frac{1}{\text{CFL}} + \frac{1}{\text{LTO}}, \qquad \text{CFL} = \frac{\text{LFL}\cdot\text{LTO}}{\text{LTO} - \text{LFL}},$$4-z1 ray loss on spherical ray parameterization, which weights rays near the image center and image edges equally on the sphere, whereas EPE is more sensitive to edge rays. The authors therefore suggest that future work may need objectives more directly aligned with EPE, especially for distortion estimation (Liang et al., 6 Jul 2026).

The benchmark is framed as infrastructure for downstream 3D vision rather than as an end in itself. The stated integration path is to predict per-frame $$\frac{1}{\text{LFL}} = \frac{1}{\text{CFL}} + \frac{1}{\text{LTO}}, \qquad \text{CFL} = \frac{\text{LFL}\cdot\text{LTO}}{\text{LTO} - \text{LFL}},$$5 and distortion from each RGB frame and then feed those estimates into existing systems such as COLMAP, DROID-SLAM, DPVO, ORB-SLAM2, NeRF, or 3D Gaussian Splatting. This suggests a modular deployment model in which intrinsics prediction serves as a front-end calibration stage for geometry and rendering pipelines (Liang et al., 6 Jul 2026).

The limitations are explicitly enumerated. Intrinsics prediction remains difficult even with the new data. Distortion estimation remains weak. There is still a synthetic-to-real gap despite the photorealistic rendering and augmentation stack. The real benchmark covers two high-quality cinema zoom lenses, not phone cameras, GoPros, fisheye lenses, or rolling shutter. The synthetic component uses a thin lens + Brown–Conrady camera model rather than all real camera behaviors. Future directions proposed in the paper include loss functions more aligned with EPE, joint intrinsics + pose + depth estimation, integration with SLAM/NeRF, broader hardware coverage, and stronger synthetic-real adaptation strategies (Liang et al., 6 Jul 2026).

In this formulation, InFlux++ occupies a specific place in the calibration literature: it is both a training corpus and an evaluation benchmark for dynamic intrinsics, with emphasis on per-frame supervision, realistic optical variation, and video diversity. Its contribution is not a new estimator architecture, but a data-and-evaluation framework meant to make RGB-based intrinsics prediction measurable, comparable, and practically useful for dynamic-camera video (Liang et al., 6 Jul 2026).