princeton-vl/InFlux-Real

InFlux-Real

InFlux-Real is the unified real-world benchmark release for the InFlux project. It combines the original InFlux benchmark with the newly captured InFlux++ Real extension, providing per-frame ground truth camera intrinsics for videos with dynamic intrinsics.

Together, the two benchmark partitions contain 657,648 annotated frames from 720 high-resolution indoor and outdoor videos. The videos span diverse scenes, camera motions, changes in intrinsics, and dynamic objects.

Naming note: InFlux-Real refers to this complete Hugging Face repository. InFlux++ Real refers specifically to the newly captured extension introduced in the InFlux++ paper.

Project Website · Download and Extraction Guide · Submission Instructions · Live Leaderboard · Synthetic Training Data

Benchmark Partitions

InFlux-Real is organized into two top-level partitions:

• —`influx/` contains the original benchmark introduced in InFlux at NeurIPS 2025.
• —`influx_pp_real/` contains InFlux++ Real, the benchmark extension introduced in InFlux++ at ECCV 2026.

Each partition is further divided into validation and test splits. Public ground truth intrinsics are provided for the validation split. Follow the official submission instructions to evaluate on the test split.

File Structure

InFlux-Real/
├── influx/
│   ├── videos/                                  # 386 .mp4 files
│   ├── video_frame_count_and_split_v1.json
│   └── gt_validation_dict_v1.json
└── influx_pp_real/
    ├── videos/                                  # 334 .mp4 files
    ├── video_frame_count_and_split_v2.json
    └── gt_validation_dict_v2.json

After TIFF Extraction

When selected partitions are decoded using the official extraction utility, each selected partition additionally contains a frames/ directory:

<output-dir>/
├── influx/
│   ├── videos/
│   │   └── *.mp4
│   ├── frames/
│   │   └── <video-name>/
│   │       └── *.tiff
│   ├── video_frame_count_and_split_v1.json
│   └── gt_validation_dict_v1.json
└── influx_pp_real/
    ├── videos/
    │   └── *.mp4
    ├── frames/
    │   └── <video-name>/
    │       └── *.tiff
    ├── video_frame_count_and_split_v2.json
    └── gt_validation_dict_v2.json

The downloaded .mp4 files and benchmark JSON files are retained after TIFF extraction. Only selected partitions are present when downloading or extracting a subset of the release.

Downloading and Extracting the Dataset

InFlux-Real is distributed as compressed .mp4 video files. Because the complete benchmark is large, we provide utilities in the official InFlux repository for downloading the dataset from Hugging Face and decoding the videos into per-frame .tiff images.

The utilities support downloading either benchmark partition independently or downloading the complete InFlux-Real release.

Approximate storage requirements are shown below:

The compressed-video column reports the approximate size of the .mp4 files as distributed through Hugging Face. The decoded-frame column reports the approximate disk space required after converting all videos in the corresponding partition into .tiff frames.

Storage requirements may vary slightly depending on filesystem overhead and extraction settings. If retaining both the downloaded videos and the decoded frames, ensure that sufficient space is available for both representations.

Users who need only one benchmark partition may download and decode influx/ or influx_pp_real/ independently. See the official download and extraction guide for the latest commands, custom output-directory usage, and extraction details.

After extracting TIFF frames, influx-verify-real can generate a count-based completeness report by comparing the extracted frame directories with the benchmark split manifests. This report does not inspect pixel contents.

Video Files

The .mp4 files contain videos featuring dynamic camera intrinsics, camera motion, and dynamic objects. The videos are encoded using YUV 4:4:4 chroma sampling.

This format may not be supported by every browser or default system video player. For local playback, we recommend using VLC.

JSON Schemas

Frame Counts and Benchmark Splits

Files:

• —influx/video_frame_count_and_split_v1.json
• —influx_pp_real/video_frame_count_and_split_v2.json

Each top-level key is a video identifier matching the stem of its corresponding .mp4 filename. For example, bike_shot11 corresponds to bike_shot11.mp4.

{
  "bike_shot11": {
    "frame_count": 338,
    "split": "val"
  }
}

Validation Ground Truth

Files:

• —influx/gt_validation_dict_v1.json
• —influx_pp_real/gt_validation_dict_v2.json

Each top-level key is a video identifier matching the stem of its corresponding .mp4 filename. Only validation videos and their annotations are included in these files. Frames are keyed by their frame indices, represented as strings.

The numeric values below are placeholders illustrating the schema:

{
  "bike_shot11": {
    "0": {
      "intrinsics_gt": {
        "fx": 0.0,
        "fy": 0.0,
        "cx": 0.0,
        "cy": 0.0,
        "k1": 0.0,
        "k2": 0.0,
        "p1": 0.0,
        "p2": 0.0
      },
      "intrinsics_gt_extrapolated": {
        "fx": 0.0,
        "fy": 0.0,
        "cx": 0.0,
        "cy": 0.0,
        "k1": 0.0,
        "k2": 0.0,
        "p1": 0.0,
        "p2": 0.0
      },
      "lens_metadata": {
        "focal_length_mm": 0.0,
        "focus_distance_m": 0.0
      }
    }
  }
}

Per-frame Annotation Fields

intrinsics_gt

Ground truth camera intrinsics obtained through interpolation of a calibration lookup table (LUT). For frames whose lens metadata falls outside the calibrated LUT bounds, the reported intrinsics values are replaced with NaN.

The parameters follow the radial-tangential Brown–Conrady distortion model:

intrinsics_gt_extrapolated

Contains the same fields as intrinsics_gt, but also provides extrapolated intrinsics for frames whose lens metadata falls outside the calibrated LUT bounds.

lens_metadata

Contains the physical lens metadata used to query the LUTs to obtain per-frame intrinsics:

Test Set Evaluation and Submission

Ground truth annotations for the validation splits are included in this release, while ground truth for the test splits is withheld.

To evaluate a model on the test split of either benchmark partition, generate predictions in the required submission format and follow the official submission instructions. The guide describes the evaluation protocol, required prediction format, and submission procedure.

Results are private by default. After evaluation, the submitter can publish a result to the live InFlux leaderboard.

For model training or finetuning, see InFlux++ Synth, our large-scale synthetic dataset with per-frame ground truth camera intrinsics.

Citation

Please cite the papers associated with the data you use:

• —If you use only the original `influx/` partition, cite the InFlux paper.
• —If you use only the `influx_pp_real/` extension, cite the InFlux++ paper.
• —If you use the complete InFlux-Real repository or report results on both partitions, cite both papers.

InFlux

InFlux: A Benchmark for Self-Calibration of Dynamic Intrinsics of Video Cameras Neural Information Processing Systems, Datasets and Benchmarks Track, 2025

NeurIPS paper · arXiv

@inproceedings{liang2025influx,
    author = {Liang, Erich and Bhattacharjee, Roma and Dey, Sreemanti and Moschopoulos, Rafael and Wang, Caitlin and Liao, Michel and Tan, Grace and Wang, Andrew and Kayan, Karhan and Alexandropoulos, Stamatis and Deng, Jia},
    booktitle = {Advances in Neural Information Processing Systems},
    editor = {D. Belgrave and C. Zhang and H. Lin and R. Pascanu and P. Koniusz and M. Ghassemi and N. Chen},
    pages = {},
    publisher = {Curran Associates, Inc.},
    title = {{InFlux}: A Benchmark for Self-Calibration of Dynamic Intrinsics of Video Cameras},
    url = {https://proceedings.neurips.cc/paper_files/paper/2025/file/8a8eca190088852067b4e8cc1b907122-Paper-Datasets_and_Benchmarks_Track.pdf},
    volume = {38},
    year = {2025}
}

InFlux++

InFlux++: Real and Synthetic Data for Estimating Dynamic Camera Intrinsics European Conference on Computer Vision, 2026

arXiv

@inproceedings{liang2026influxpp,
    author={Liang, Erich and Kha-Uong, Caleb and Saran, Chinmaya and Dey, Sreemanti and Liu, David W. and Ouyang, Junhan and Zhou, Benjamin and Deng, Jia},
    editor={Favaro, Paolo and Kukelova, Zuzana and Maki, Atsuto and Rohrbach, Anna and Schindler, Konrad and Tombari, Federico},
    title={{InFlux++}: Real and Synthetic Data for Estimating Dynamic Camera Intrinsics},
    booktitle={Computer Vision -- ECCV 2026},
    year={2026},
    publisher={Springer Nature Switzerland},
    address={Cham},
    pages={443--460},
    isbn={978-3-032-37235-2}
}