karthik789338/picd-video-repro

PICD-Video Reproducibility Data

This repository contains processed reproducibility data, configuration files, result summaries, and paper tables for the PICD-Video project.

PICD-Video studies physics-informed dynamic causal discovery from physical video/object-state sequences. The method learns lagged dynamic interaction graphs and uses them for rollout prediction, graph recovery, and counterfactual intervention evaluation.

Contents

• data/processed/

  • Processed synthetic/oracle and visual-state datasets used for the main experiments.
  • Includes oracle states, smoothed keypoint states, rendered detector states, CNN detector states, heatmap detector states, and SAVi-hinted diagnostic states when available.

• configs/

  • YAML configs used for training and evaluation.

• paper_tables/

  • CSV tables used to build the manuscript tables.

• paper_assets/

  • LaTeX tables and result summaries.

• results/summaries/

  • JSON/CSV evaluation summaries copied from experiment output folders.

Not included

This first reproducibility release does not include raw CLEVRER or CoPhy videos/data. Those will be released separately as derived/preprocessed benchmark-specific repositories when licensing and attribution are finalized.

This repository also does not include model checkpoints by default. Checkpoints can be regenerated using the public GitHub code and configs.

Code

GitHub repository:

https://github.com/karthik789338/picd-video

Intended use

This dataset is intended for:

• reproducing manuscript tables,
• checking evaluation summaries,
• rerunning PICD-Video experiments from processed states,
• comparing dynamic graph, static graph, no-graph, and recurrent baseline settings.

Limitations

Some external benchmark results are diagnostic stress tests rather than main claims. SAVi-hinted variants are included as failed/diagnostic perception ablations and should not be interpreted as successful object-centric video encoders.