README.md

---
license: cc-by-4.0
library_name: pytorch
tags:
  - digital-twins
  - scene-graphs
  - dynamic-scene-graphs
  - robotics
  - support-relations
  - graph-maintenance
  - counterfactual-evidence
  - pytorch
---

# CF-SupportNet

CF-SupportNet is the learned edge-scoring component of **MedPhyGraph** (ECCV 2026 TwinWorld Workshop). It is **not** the full MedPhyGraph framework by itself: MedPhyGraph additionally applies deterministic State Consistency and Union-Based Transition-Aware Consistency — non-learned, code-only operations — on top of these scores to produce the final maintained support graph. See the [project page](https://medphygraph.github.io/) for the paper and complete pipeline code.

## Files

| File | Role |
|---|---|
| `health_dyphygraph_r1.0_seed0.pt` | Primary checkpoint — used for all headline results in the paper (ρ=1.0, seed 0) |
| `health_dyphygraph_r1.0_seed{1-4}.pt` | Used only for the paper's multi-seed reproducibility audit; not intended as alternative production checkpoints |

## Architecture

- One-layer GRU, hidden width 64 (temporal encoder over counterfactual evidence channels)
- Geometry branch: 2 → 32 → 32 MLP (static geometric channels)
- Fusion branch: 96 → 64 → 64 MLP, dropout 0.1
- **25,409 trainable parameters** total

Full architectural and training details are in the paper, Section 3.4.

## Results

Frozen seed-0 checkpoint, full observation (ρ=1.0):

| Protocol | Transition-Macro Dyn-F1 | Pooled Delta Micro-F1 | Transfer Dyn-F1 |
|---|---|---|---|
| Core (15 transfers, Isaac + Procedural) | 1.000 | 1.000 | 1.000 |
| Expanded (217 transfers, 19 templates) | — | — | 0.998 |

Across seeds 0–4, pooled Transfer Dyn-F1 on the expanded suite is 0.997 ± 0.001.

## Loading the checkpoint

```python
from huggingface_hub import hf_hub_download
import torch

path = hf_hub_download(
    repo_id="MedPhyGraph/CF-SupportNet",
    filename="health_dyphygraph_r1.0_seed0.pt",
)
state_dict = torch.load(path, map_location="cpu", weights_only=True)
```

> This repository hosts raw PyTorch weights only, not a packaged `from_pretrained`-compatible class. To reconstruct the model, define the architecture above (or import it from the [code repository](https://medphygraph.github.io/)) and load `state_dict` into it directly.

## Limitations

- The counterfactual evidence used at both training and inference comes from an analytic AABB-based host-removal proxy, not a full physics simulator — it does not model mesh collisions, contact forces, or materials.
- Operates downstream of perception on structured object states (poses, categories, AABB extents); it does not take raw RGB-D or perception input.
- Evaluates direct, primary `SupportedBy` relations only.
- Assumes the correct destination candidate is present in the candidate set at inference; if it is not, a support transfer cannot be recovered by the scorer or the downstream graph-maintenance stages.
- Does not model distributed support, articulated contact, or temporary/transient support.

## Related artifacts

- Structured training/evaluation data (procedural + Isaac for Healthcare-derived scenes): `MedPhyGraph/support-graph-data` *(forthcoming)*

## Citation

```bibtex
@inproceedings{gholizadeh2026medphygraph,
  title     = {MedPhyGraph: Counterfactual Support-Graph Maintenance for Dynamic Built-Environment Digital Twins},
  author    = {Gholizadeh HamlAbadi, Kamran and Vahdati, Monica and El Saddik, Abdulmotaleb},
  booktitle = {ECCV 2026 Workshop on Visual Intelligence for Built Environment Digital Twins (TwinWorld)},
  year      = {2026}
}
```