CoolFace
Datasetpublic

YuxinJ/camera-motion-eval-assets

Camera Motion Evaluation Assets This repository contains generated video samples and derived camera-motion diagnostics from a sequence of development snapshots. It is intended for qualitative comparison and evaluator reproducibility. Layout training/: short development samples. Each video includes the generated frames and a synchronized trajectory panel. benchmark/: fixed long-form rollouts, numeric evaluator outputs, recovered pose trajectories, and side-by-side… See the full description on the dataset page: https://huggingface.co/datasets/YuxinJ/camera-motion-eval-assets.

sourceHugging Faceotherupdated 26d agoView on Hugging Face
0likes83downloads
Dataset Card

Camera Motion Evaluation Assets

This repository contains generated video samples and derived camera-motion diagnostics from a sequence of development snapshots. It is intended for qualitative comparison and evaluator reproducibility.

Layout

  • training/: short development samples. Each video includes the generated frames and a synchronized trajectory panel.
  • benchmark/: fixed long-form rollouts, numeric evaluator outputs, recovered pose trajectories, and side-by-side trajectory renders.
  • analysis/: compact per-scene comparison tables derived from the published evaluator outputs.

The labels series-main, series-a, and series-b identify separate data mixtures. Snapshot labels are relative development indices and should not be interpreted as release versions.

Conventions

  • Videos are 16 fps.
  • The clean benchmark videos are kept separate from annotated renders so that overlays cannot contaminate numeric evaluation.
  • Camera errors are reported after per-scene similarity alignment. Fitted scale is therefore a diagnostic rather than a deployment calibration constant.
  • Files under metrics/ are evaluator outputs; files under trajectories/ are derived pose estimates and references used by those metrics.

Scope

Only generated media and derived evaluation artifacts are included. Model weights, optimizer state, logs, source media, and training records are not part of this repository. Upstream benchmark terms continue to apply to benchmark inputs and derived outputs.