---
license: cc-by-4.0
pretty_name: PICD-Video CoPhy CollisionCF_4 Processed States
tags:
- computer-vision
- causal-discovery
- physical-reasoning
- counterfactual-reasoning
- cophy
- collision-dynamics
- graph-neural-networks
---

# PICD-Video CoPhy CollisionCF_4 Processed States

This dataset contains the processed CoPhy CollisionCF_4 state-level data used for external diagnostic evaluation in the PICD-Video project.

PICD-Video studies physics-informed dynamic causal discovery from physical video/object-state sequences. This processed dataset supports external evaluation of collision-style dynamics, sparse event graph recovery, and counterfactual rollout diagnostics.

## Contents

- `data/processed/cophy/collisionCF_4_v2/`
  - `train.npz`
  - `val.npz`
  - `test.npz`
  - `counterfactual_test.npz`
  - `metadata.json`
  - `manifest.csv`

- `configs/cophy/`
  - YAML config used for model training/evaluation.

- `results/`
  - Evaluation outputs and summary CSV files.

- `paper_assets/`
  - Paper-ready supplementary table files when available.

## Preprocessing

The original CoPhy state arrays were converted into PICD-compatible object states:

`[x, y, vx, vy, mass, radius]`

Sparse event-style pseudo-edges were generated using a distance radius, nearest-neighbor filtering, and an approaching-motion criterion. The final processed CollisionCF_4 setup used 4 objects and 15 timesteps.

## Important note

This repository contains processed state-level derivative data and evaluation summaries. It does not include raw CoPhy videos. Users should refer to the original CoPhy project for the source benchmark.

## Code

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

## Intended use

- Reproducing PICD-Video CoPhy CollisionCF_4 diagnostic results.
- Evaluating physical collision dynamics from processed object states.
- Comparing graph-based, no-graph, static-graph, and trajectory-only baselines.

## Limitations

In PICD-Video experiments, CoPhy CollisionCF_4 is used as an external diagnostic benchmark. No-graph and CPM-style models are strong trajectory baselines on this processed route, while PICD provides dynamic graph/event structure.