Data

---

Our dataset is available on Hugging Face.

Hugging Face

Please follow the instructions on the Hugging Face repository and the code repository to download and use the dataset.

Instructions on how to use the dataset are given below.

Dataset Structure

The dataset includes 386 mp4 videos and 2 json files in the root directory:

• Dynamic Intrinsics Videos (386 mp4 files): Videos with dynamic intrinsics and other arbitrary camera motions.
• Frame counts and Val-Test Split (1 JSON file: video_frame_count_and_split.json): JSON file mapping video names to the number of frames in the mp4 and whether this video is in the validation or test split.
• Validation Ground Truth (1 JSON file: gt_validation_dict.json): JSON file mapping validation video frames to ground truth intrinsics values.

JSON File Structures

• video_frame_count_and_split.json:
  • Video name (e.g. video_shot_1)
  • frame_count – How many frames in the mp4.
  • split – Does this video belong in the validation or test split.
• gt_validation_dict.json:
  • Video name (e.g. video_shot_1)
  • Frame number (e.g. 0)
  • intrinsics_gt – Ground truth intrinsics interpolated from the Look-Up Table (LUT).
  • fx
  • fy
  • cx
  • cy
  • k1
  • k2
  • p1
  • p2
  • intrinsics_gt_extrapolated – Intrinsics extrapolated from the LUT, used if lens metadata exceeds LUT bounds.
  • [Same as intrinsics_gt]
  • lens_metadata – Physical lens values.
  • focal_length_mm
  • focus_distance_m
  • Frame number (e.g. 1)
  • ...

For any questions or assistance, please contact us at: influxbenchmark@gmail.com