Redirect4D-Bench

Redirect4D-Bench is a benchmark for evaluating camera redirection on monocular dynamic videos.

For documentation, the construction pipeline, evaluation scripts, and usage examples, please refer to the GitHub repository:

https://github.com/VVeiCao/redirect4d-bench

Quick start: sample download and preview

Download the public sample (~2 GB):

hf download vveicao/redirect4d-bench \
  --repo-type dataset \
  --include 'sample/**' \
  --local-dir data/redirect4d_bench

Preview the sample with the Viser viewer from the GitHub repository:

python scripts/visualization/serve_pointcloud_viser.py \
  --dataset-root data/redirect4d_bench/sample \
  --port 8091

Forward port 8091 and open it in your browser. The viewer shows the animated foreground/background 4D point clouds, the source RGB video and source mask, the active target trajectory and camera, and the target mask and depth videos for the active trajectory.

The sample is a representative subset of the public release: two complete tracks (bear_NnAlfavy2us_003_001_seq1 and elephant_4F0hzklQejU_010_001_seq1) selected from the 83 retained cases to illustrate the full per-case bundle. Unlike the full release, the sample also ships with source RGB video and frames, so the preview works end-to-end without recovering source clips locally.

Full dataset

Download the full benchmark assets (~54 GB):

hf download vveicao/redirect4d-bench \
  --repo-type dataset \
  --local-dir data/redirect4d_bench

The public full dataset does not redistribute source RGB videos or frames; the sample/ folder contains the small subset that includes source RGB for quick preview. Metadata carries the YouTube ids, crop boxes, frame ids, fps, camera intrinsics, and trajectory names needed to reconstruct source clips locally.

Source RGB recovery

The download script in the GitHub repository reconstructs the original source clips from the YouTube ids and metadata in this dataset. If a video is region-blocked, geo-fenced, or has been removed and the script cannot recover it, please reach out to the maintainers via the contact information listed on the GitHub repository.

Layout

data/redirect4d_bench/
  metadata.json
  tracks.jsonl
  cases.jsonl
  sample/                          # small public sample (includes source RGB)
  tracks/<track>/
    camera.json
    masks/00000.png ... 00044.png
    mask_video.mp4
    pointcloud/
      global_background.ply
      <frame>/foreground_1_view.{npz,ply}
      <frame>/foreground_5_views_aligned_smooth.{npz,ply}
    redirected/<trajectory>/
      trajectory.json
      mask.mp4                     # pseudo-GT target mask
      depth.mp4                    # pseudo-GT target depth
      prompt.txt

License

CC-BY-NC-4.0.